CN118193853B - A method, system, device and storage medium for recommending points of interest for random groups - Google Patents

Abstract

The application relates to the technical field of business field data prediction recommendation, in particular to a random group interest point recommendation method, a random group interest point recommendation system, a random group interest point recommendation device and a random group interest point recommendation storage medium; in order to solve the problems that in the prior art, the random group interest point prediction result is low, the recommendation accuracy is low and the commercial experience sense of a user is influenced, the method comprises the steps of firstly combining the feature of the interest point set to be recommended obtained based on similar users of the random group with the influence degree of the character of the user to obtain a fitting feature representation of the random group; then, performing multi-layer graph neural network processing on the interest interaction structure graph to obtain feature representation of each interest point to be recommended; then, probability mapping processing is carried out on the feature representation of the interest points to be recommended, and the preferred interest points are obtained; finally, after the expected value of the random user is obtained according to the prediction scores of the preferred interest points, multi-negotiation recommendation processing is carried out to obtain the optimal recommended interest points; the method and the device are applied to the field of business data prediction, and can improve the recommendation accuracy and the business experience of the user.

Description

A method, system, device and storage medium for recommending points of interest for random groups

Technical Field

The present invention relates to the technical field of commercial data prediction and recommendation, and in particular to a method, system, device and storage medium for recommending points of interest of a random group.

Background technique

With the advent of the big data era, the amount of information generated every minute and every second is growing exponentially. Faced with a large amount of data, relevant commercial software is difficult to quickly predict accurate services and recommend them to target users, which can easily reduce user experience and commercial content consumption. Group point of interest recommendation is one of the key technologies to overcome the problem of data overload. By deeply analyzing the huge amount of user check-in information, it recommends appropriate points of interest to the group in the massive data, effectively reducing the interference caused by information overload.

There are two types of groups in society: fixed groups and random groups. Fixed groups are the most common type of groups in society today, such as families, interest groups, etc. There are often strong social relationships between members of fixed groups and the differences in preferences between members are relatively small. At the same time, fixed groups have a long life cycle and will not disappear due to the end of the recommendation task. Random groups refer to people who gather together at a specific moment for a common purpose or to participate in an activity. There is usually no social or other type of relationship between members of a random group. Once the purpose is achieved or the activity ends, the group will disband.

However, existing work mainly provides POI recommendations for fixed groups. Due to the differences between random groups and fixed groups, the existing POI recommendation methods for fixed groups are difficult to apply to random groups. Moreover, the recommendation results in the current group recommendation research have not been discussed and negotiated based on the personalities of group users, resulting in low POI prediction results and low recommendation accuracy, which can easily reduce users' commercial experience and affect commercial content consumption.

Summary of the invention

The purpose of the present invention is to provide a method, system, device and storage medium for recommending points of interest in random groups, which have high prediction results of points of interest, high recommendation accuracy and can enhance the user's business experience.

The technical solution of the present invention is as follows:

A method for recommending points of interest of a random group includes the following operations:

S1. Obtain similar users of each random user in the random group, and all similar users form a similar user set; the product of the user feature vector of the similar user and the feature vector of the random user is greater than the similarity threshold; obtain the points of interest that each similar user has interacted with in the similar user set as the points of interest to be recommended; all the points of interest to be recommended form a set of points of interest to be recommended;

S2. Based on the set of interest points to be recommended, the relative influence weight of each random user in the random group is obtained; based on the relative influence weight of each random user and the influence of each user's personality, the absolute influence weight of each random user is obtained; the absolute influence weight of each random user is multiplied by the feature vector of each random user, and then aggregated to obtain the fitting feature representation of the random group;

S3. Based on the recommended interest points, construct an interest interaction structure diagram for each recommended interest point; the interaction structure diagram is composed of the recommended interest points and/or similar users as nodes, and the interaction edges between the recommended interest points and similar users; the interest interaction structure diagram of each recommended interest point is processed by a multi-layer graph neural network to obtain the feature representation of each recommended interest point; the feature representations of all the recommended interest points form a feature representation set of the recommended interest points; in the feature representation set of the recommended interest points, each feature representation of the recommended interest point is respectively matched with the random group fitting feature representation, and a predicted score is obtained by probability mapping; the recommended interest points corresponding to the predicted score greater than the score threshold are taken as the preferred interest points; all the preferred interest points form a preferred interest point set;

S4. Based on the predicted score of each preferred point of interest in the preferred point of interest set, the expected value of each random user for each preferred point of interest is obtained; based on the expected value of each random user for each preferred point of interest, a multi-negotiation recommendation process is performed to obtain the optimal recommended point of interest for recommendation to the random group.

The specific operations of multi-negotiated recommendation processing in S4 are as follows: S4.1. In the random group, each random user takes the preferred interest point corresponding to the maximum expected value of their respective users as their respective recommended items; all recommended items form a recommended item set; S4.2. Determine whether there is a corresponding recommended item in the recommended item set whose expected value is not less than the expected value of other recommended items of the same user; if so, take the corresponding recommended item with an expected value not less than the expected value of other recommended items of the same user as the optimal recommended interest point for recommendation to the random group; if not, execute S4.3; S4.3. Based on the recommended item set and the corresponding expected value, obtain the risk-taking willingness of each random user; take the random user with the minimum risk-taking willingness as the target user; select a preferred interest point corresponding to a target user whose expected value is not the maximum as the target recommended item; replace the recommended item of the target user in the recommended item set with the target recommended item to obtain a negotiated recommended item set for executing the operation of S4.2.

In S2, the method for obtaining the relative influence weight of the random user is as follows: obtain the differential vectors of the random user to be processed and other random users in the random group, and obtain the fused vector to be processed after all the differential vectors are fused; obtain the fused vector to be processed, and multiply it by the feature vector of the current POI to be recommended in the POI set to be recommended to obtain the current comparison vector to be processed; all the comparison vectors to be processed are processed by concatenation, nonlinearity and activation function in turn to obtain the relative influence weight of the random user to be processed.

In S2, the method for obtaining the absolute influence weight of a random user is as follows: a random user is randomly selected from the random group as the representative random user; the absolute value of the difference between the relative influence weights of other random users and the relative influence weight of the representative random user is used as the absolute factors of other random users; the absolute factors of other random users are multiplied by their respective user personality influences to obtain the absolute influence weights of other random users; the relative influence weight of the representative random user is multiplied by the corresponding user personality influence to obtain the absolute influence weight of the representative random user.

The method for obtaining the user's personality influence is as follows: multiply the user's score under the Big Five personality model by the corresponding personality influence ratio, perform summation and normalization, and obtain the user's personality influence.

In S4, the method for obtaining the expected value is: if the random user has evaluated the preferred interest point, the score of the random user on the preferred interest point is used as the expected value; if the random user has not evaluated the preferred interest point, the predicted score corresponding to the preferred interest point is used as the expected value.

In S4, the method for obtaining the risk-taking willingness is as follows: if the expected value of the random user for the recommended item set is 0, the risk-taking willingness of the random user is 1; if the expected value of the random user for the recommended item set is not 0, the ratio of the difference between the expected average value of the random user and the minimum expected value in the recommended item set to the expected average value of the random user is taken as the risk-taking willingness of the random user.

A random group point of interest recommendation system, comprising:

The module for generating the set of interest points to be recommended is used to obtain similar users of each random user in the random group, and all similar users form a similar user set; the product of the user feature vector of the similar user and the feature vector of the random user is greater than the similarity threshold; the interest points that each similar user has interacted with in the similar user set are obtained as the interest points to be recommended; all the interest points to be recommended form a set of interest points to be recommended;

The random group fitting feature representation generation module is used to obtain the relative influence weight of each random user in the random group based on the set of interest points to be recommended; based on the relative influence weight of each random user and the influence of each user's personality, the absolute influence weight of each random user is obtained; the absolute influence weight of each random user is multiplied by the respective random user feature vector, and then aggregated to obtain the random group fitting feature representation;

The preferred interest point set generation module is used to construct an interest interaction structure diagram for each recommended interest point based on the interest points to be recommended; the interaction structure diagram is composed of the interest points to be recommended and/or similar users as nodes, and the interaction edges between the interest points to be recommended and similar users; the interest interaction structure diagram of each recommended interest point is processed by a multi-layer graph neural network to obtain the feature representation of each interest point to be recommended; the feature representations of all the interest points to be recommended form a feature representation set of the interest points to be recommended; in the feature representation set of the interest points to be recommended, the feature representation of each interest point to be recommended is respectively matched with the feature representation of the random group fitting, and a predicted score is obtained by probability mapping; the interest points to be recommended corresponding to the predicted score greater than the score threshold are taken as the preferred interest points; all the preferred interest points form a preferred interest point set;

The multi-negotiation recommendation processing and interest point recommendation module is used to obtain the expected value of each random user for each preferred interest point based on the predicted score of each preferred interest point in the preferred interest point set; based on the expected value of each random user for each preferred interest point, multi-negotiation recommendation processing is performed to obtain the optimal recommended interest point for recommendation to the random group.

A device for recommending points of interest for a random group includes a processor and a memory, wherein the processor implements the above-mentioned method for recommending points of interest for a random group when executing a computer program stored in the memory.

A computer-readable storage medium is used to store a computer program, wherein when the computer program is executed by a processor, the above-mentioned method for recommending points of interest of a random group is implemented.

The beneficial effects of the present invention are:

The present invention provides a method for recommending points of interest for a random group. First, based on the size relationship between the product of feature vectors and a similarity threshold, similar users with similar preferences as random users in the random group are selected, and points of interest that similar users have interacted with are obtained as points of interest to be recommended that the random group may be interested in, thereby narrowing the scope of selecting points of interest; then, according to the characteristics of the set of points of interest to be recommended, the relative influence weight of each random user is obtained, and the relative influence weight of each random user is combined with the user personality influence that can reflect the respective personality characteristics to obtain the absolute influence weight of each random user in the random group; after multiplying the absolute influence weight of each random user with the respective random user feature vector, aggregation processing is performed to obtain a random group fitting feature representation that can reflect the overall preference of the random group; then, each point of interest to be recommended and the similar users who have interacted with it are constructed into a structural graph, and a multi-layer graph neural network is performed. After network processing, the message propagation between the recommended interest points and similar users is realized, and the feature representation of each interest point to be recommended with rich features and stronger expression ability is obtained; then, the feature representation of each interest point to be recommended is probability mapped with the random group fitting feature representation to obtain a predicted score; and the recommended interest points with predicted scores greater than the score threshold are used as preferred interest points that are more easily accepted by the random group, further narrowing the range of selected interest points; finally, according to the predicted scores of the preferred interest points, the expected values of each random user for the preferred interest points are obtained, and based on the expected values of each random user, multi-negotiation recommendation processing is performed on the preferred interest points to obtain the optimal recommended interest points that satisfy all random users for recommendation to the random group; this method screens and negotiates interest points based on the user's personality, and the interest point prediction results are more accurate, the recommendation accuracy is higher, and the user's commercial experience can be improved.

BRIEF DESCRIPTION OF THE DRAWINGS

By reading the detailed description of the preferred embodiment below, the scheme and advantages of the present application will become clear to those skilled in the art. The accompanying drawings are only for the purpose of illustrating the preferred embodiment and are not to be considered as limiting the present invention.

In the attached picture:

FIG1 is a schematic diagram of a flow chart of a method of the present embodiment;

FIG2 is a graph showing the experimental results of the effect of group size on group owner satisfaction with multi-negotiation processing in the method of this embodiment on a Foursquare dataset in an embodiment;

FIG3 is a graph showing the experimental results of the effect of group size on the dispersion of member satisfaction in the multi-negotiation process in the method of the present embodiment on the Foursquare dataset in the embodiment;

FIG4 is a graph showing experimental comparison results of the effect of group size on group owner satisfaction with negotiation processing in the method of this embodiment and the random selection method on the Foursquare dataset in the embodiment.

Detailed ways

Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings.

This embodiment provides a method for recommending points of interest in a random group, referring to FIG1 , including the following operations:

S1. Obtain similar users of each random user in the random group, and all similar users form a similar user set; the product of the user feature vector of the similar user and the feature vector of the random user is greater than the similarity threshold; obtain the points of interest that each similar user has interacted with in the similar user set as the points of interest to be recommended; all the points of interest to be recommended form a set of points of interest to be recommended;

S2. Based on the set of interest points to be recommended, the relative influence weight of each random user in the random group is obtained; based on the relative influence weight of each random user and the influence of each user's personality, the absolute influence weight of each random user is obtained; the absolute influence weight of each random user is multiplied by the feature vector of each random user, and then aggregated to obtain a random group fitting feature representation;

S3. Based on the interest points to be recommended, construct an interest interaction structure diagram for each recommended interest point; the interaction structure diagram is composed of the interest points to be recommended and/or similar users as nodes, and the interaction edges between the interest points to be recommended and similar users; the interest interaction structure diagram of each recommended interest point is processed by a multi-layer graph neural network to obtain a feature representation of each interest point to be recommended; all the feature representations of the interest points to be recommended form a feature representation set of the interest points to be recommended; in the feature representation set of the interest points to be recommended, each feature representation of the interest points to be recommended is respectively matched with the random group fitting feature representation, and a predicted score is obtained by probability mapping; the interest points to be recommended corresponding to the predicted score greater than the score threshold are taken as preferred interest points; all the preferred interest points form a preferred interest point set;

S4. Based on the predicted score of each preferred point of interest in the preferred point of interest set, obtain the expected value of each random user for each preferred point of interest; based on the expected value of each random user for each preferred point of interest, perform multi-negotiation recommendation processing to obtain the optimal recommended point of interest for recommendation to the random group.

S1. Obtain similar users of each random user in the random group, and all similar users form a similar user set; the product of the user feature vector of the similar user and the feature vector of the random user is greater than the similarity threshold; obtain the points of interest that each similar user has interacted with in the similar user set as the points of interest to be recommended; all the points of interest to be recommended form a set of points of interest to be recommended.

Based on the relationship between the product of the feature vectors and the similarity threshold, similar users with similar preferences as random users in the random group are selected, and the points of interest that similar users have interacted with are obtained as the recommended points of interest that the random group may be interested in. This narrows the scope of selecting points of interest and improves the recommendation efficiency while improving the accuracy of the recommendation results.

First, the corresponding users whose product of the user feature vector and the random user feature vector is greater than the similarity threshold are regarded as similar users. For example, if the product of the user feature vector of user A and the random user feature vector of random user B is greater than the similarity threshold, user A is a similar user of random user B. The user feature vector is obtained by concatenating the user's identity information and interaction information and embedding them. All similar users form a similar user set with similar preferences to the random group.

Then, each interest point that has interacted with any similar user in the similar user set, that is, obtaining the interest points that each similar user in the similar user set has interacted with, is taken as an interest point that the random group may like, that is, as the interest point to be recommended. All the interest points to be recommended form a set of interest points to be recommended.

S2. Based on the set of interest points to be recommended, the relative influence weight of each random user in the random group is obtained; based on the relative influence weight of each random user and the influence of each user's personality, the absolute influence weight of each random user is obtained; the absolute influence weight of each random user is multiplied by the feature vector of each random user, and then aggregated to obtain the fitting feature representation of the random group.

According to the characteristics of the set of interest points to be recommended, the relative influence weight of each random user is obtained, and the relative influence weight of each random user is combined with the user personality influence that can reflect their respective personality characteristics to obtain the absolute influence weight of each random user in the random group; by multiplying the absolute influence weight of each random user with the respective random user feature vector and performing aggregation processing, a random group fitting feature representation that can reflect the overall preference of the random group is obtained.

First, based on the set of points of interest to be recommended, the relative influence weight of each random user in the random group is obtained. The operation of obtaining the relative influence weight of the random user is as follows: the differential vectors of the random user to be processed and other random users in the random group are obtained, and all differential vectors are fused to obtain the fused vector to be processed; the fused vector to be processed is obtained, and the product of the feature vector of the current point of interest to be recommended in the set of points of interest to be recommended is obtained to obtain the current comparison vector to be processed; all comparison vectors to be processed are processed by concatenation, nonlinear processing (which can be achieved by Relu activation function) and activation function (which can be achieved by Softmax activation function) in turn to obtain the relative influence weight of the random user to be processed. The feature vector of the point of interest to be recommended is obtained by concatenating the address information and interaction information of the point of interest to be recommended and embedding them.

This method compares different random users and obtains relative influence weights, which can reflect the true importance of random users in the random group.

Among them, the operation of obtaining the fusion vector to be processed can be achieved by the following formula:

,

C _m is the to-be-processed fusion vector of the to-be-processed random user m , um _{is the} to-be-processed random user feature vector of the to-be-processed random user m , un _is the random user feature vector of the random user n , and N is the total number of other random users.

Then, based on the relative influence weight of each random user and the influence of their respective user personality, the absolute influence weight of each random user is obtained. The absolute influence weight of a random user is obtained by: selecting a random user from the random group as the representative random user; taking the absolute value of the difference between the relative influence weight of other random users and the relative influence weight of the representative random user as the absolute factors of other random users; multiplying the absolute factors of other random users by their respective user personality influences to obtain the absolute influence weights of other random users; multiplying the relative influence weight of the representative random user by the corresponding user personality influence to obtain the absolute influence weight of the representative random user. Random members with different personalities often have different influences on the final decision of the random group. Group members with positive personalities may be more persuasive and determined, thus dominating the group and making the random group's choice more inclined to his preference. By considering the influence of the user personality of the random user, different influence weights can be assigned to the random group members according to their personality characteristics when calculating the random group feature fitting, so as to obtain a random group fitting feature representation that can better reflect the overall preference of the random group, which is used to improve the accuracy of subsequent interest point recommendation.

The method for obtaining the user's personality influence is as follows: multiply the user's score under the Big Five personality model by the corresponding personality influence ratio, perform summation and normalization, and obtain the user's personality influence.

The influence of user personality can be realized through the following formula:

,

Q _m is the user personality influence of random user m , MinMaxScaler() is the normalization function, [-1,1,1,-1,1] is an array composed of the five personality influence ratios in the Big Five personality model, and PER _m is an array formed by the user's score value under the Big Five personality model.

The Big Five personality model is an existing technology. The five personalities in the Big Five personality model are: Openness: Those with high scores may be more willing to try new concepts and ideas, and may be more open to the opinions put forward, but may also be more receptive to the opinions of others. Those with low scores may be more conservative and may be more persistent in their own opinions, but may also find it difficult to accept new concepts; Conscientiousness: Those with high scores may be more organized and goal-oriented, and may be more persistent in their own opinions, especially when these opinions are related to achieving common goals. Those with low scores may be more casual in decision-making and may be more flexible in their opinions; Extraversion: Those with high scores may be more willing to try new concepts and ideas, and may be more open to the opinions put forward, but may also be more receptive to the opinions of others. sion): Those with high scores may be better at socializing and may find it easier to express their opinions and defend their opinions in a group. Those with low scores may be more introverted in a group and may need more effort to express and stick to their opinions. Agreeableness: Those with high scores may be more cooperative, may be more willing to compromise and accept the opinions of others, and may be more flexible in a group. Those with low scores may be more persistent in their own opinions and may find it more difficult to compromise in a group. Emotional stability (Neuroticism): Those with high scores may remain calmer under pressure and may find it easier to stick to their opinions. Those with low scores may be more easily affected under pressure and may find it more challenging to stick to their own opinions.

Finally, the absolute influence weight of each random user is multiplied by each random user feature vector, and then the aggregation process is achieved by summing up to obtain the random group fitting feature representation.

S3. Based on the set of interest points to be recommended, construct an interest interaction structure graph for each recommended interest point; the interest interaction structure graph is composed of the interest points to be recommended and/or similar users as nodes, and the interaction edges between the interest points to be recommended and similar users; the interest interaction structure graph of each recommended interest point is processed by a multi-layer graph neural network to obtain the feature representation of each interest point to be recommended; the feature representations of all interest points to be recommended form a feature representation set of interest points to be recommended; in the feature representation set of interest points to be recommended, each feature representation of the interest point to be recommended is respectively matched with the feature representation of the random group fitting, and is processed by probability mapping to obtain a predicted score; the interest points to be recommended corresponding to the predicted scores greater than the score threshold are regarded as preferred interest points; all preferred interest points form a preferred interest point set.

By constructing a structural graph with similar users that each interest point to be recommended has interacted with, and performing multi-layer graph neural network processing, the message propagation between the interest points to be recommended and similar users is realized, and a feature representation of each interest point to be recommended with rich features and stronger expression ability is obtained; then the feature representation of each interest point to be recommended is probabilistically mapped with the random group fitting feature representation to obtain a predicted score; finally, the interest points to be recommended whose predicted scores are greater than the score threshold are used as preferred interest points that are more easily accepted by the random group, thereby further narrowing the scope of selecting interest points and improving the accuracy of the recommendation results.

First, based on the set of interest points to be recommended, an interest interaction structure graph of each recommended interest point is constructed; the interest interaction structure graph is composed of the interest points to be recommended and similar users as nodes, as well as the interaction edges between the interest points to be recommended and similar users. In the current interest interaction structure graph, the current recommended interest point and the similar users with which it has interacted are connected through interaction edges, which can more clearly express the interaction information of the interest points to be recommended. The interest interaction structure graph can be used in subsequent multi-layer graph neural network processing to propagate information from similar users who have interacted with the recommended interest points to the recommended interest points.

At the same time, based on the similar user set, a user interaction structure diagram for each similar user can be constructed; in the current user interaction structure diagram, the recommended points of interest that the current similar user has interacted with are connected through interaction edges, which can more clearly express the interaction information of similar users. The user interaction structure diagram can be used in subsequent multi-layer graph neural network processing to propagate information about the recommended points of interest that have interacted with similar users to similar users.

Then, the interest interaction structure diagram of each recommended interest point is processed by graph neural network to obtain the feature representation of each interest point to be recommended.

The process of multi-layer graph neural network processing can be as follows (multi-layer graph neural network processing is not performed on the user interaction structure graph product of similar users, that is, the recommended interest points do not propagate information to similar users): based on the interest interaction structure graph of the current recommended interest points in the previous layer, the feature representation of the current interest points to be recommended in the previous layer and the feature vector of similar users, the current layer attention coefficient is obtained; all the current layer attention coefficients form the current layer attention coefficient set; in the current layer attention coefficient set, each current layer attention coefficient is multiplied by the corresponding similar user feature vector, and then summed to obtain the current layer similar user propagation information; the current layer similar user propagation information and the feature representation of the current interest points to be recommended in the previous layer are concatenated and fully connected to obtain the feature representation of the current interest points to be recommended in the current layer.

In the process of multi-layer graph neural network processing, it is also possible to (process the user interaction structure graph product of similar users with multi-layer graph neural network, and propagate information of the recommended interest points to similar users), based on the interest interaction structure graph of the current recommended interest points of the previous layer, the feature representation of the current interest points to be recommended of the previous layer and the feature representation of similar users of the previous layer, obtain the current layer attention coefficient; all the current layer attention coefficients form the current layer attention coefficient set; in the current layer attention coefficient set, each current layer attention coefficient is multiplied by the corresponding feature representation of similar users of the previous layer, and then summed to obtain the current layer similar user propagation information; the current layer similar user propagation information and the feature representation of the current interest points to be recommended of the previous layer are spliced and fully connected to obtain the feature representation of the current interest points to be recommended of the current layer. The method for obtaining the feature representation of the current similar users of the current layer is similar to the method for obtaining the feature representation of the current interest points to be recommended of the current layer. In order to save space, it is no longer described in words and can be obtained through the following formula.

Specifically, the feature representation of the current recommended POI can be obtained by the following formula:

,

,

,

,

,

h ^p(i,l) is the feature representation of the recommended interest point of the i -th interest point to be recommended in the l- th layer; p _i is the feature vector of the recommended interest point of the i -th interest point to be recommended, u _j is the similar user feature vector of the j -th similar user, m ^l (p _i ) is the similar user propagation information of the i- th interest point to be recommended in the l- th layer, m ^l (u _j ) is the similar user feature representation of the j -th similar user in the l -1 layer, p _i =m ⁰ (p _i ) , u _j =m ⁰ (u _j ); MLP( ) is the fully connected processing, W ₁ , W ₂ are the first learnable weight and the second learnable weight respectively, dropout( ) is the parameter loss processing, b ₁ , b ₂ are the first learnable bias parameter and the second learnable bias parameter respectively, α ^ij,l-1 is the attention coefficient between the i -th interest point to be recommended and the j -th similar user in the l-1-th layer, m ^l-1 (u _j ) is the similar user feature representation of the j -th similar user in the l-1-th layer, α ^jk,l-2 is the attention coefficient between the k -th interest point to be recommended in the l-2th layer and the j -th similar user, and m ^l-2 (p _k ) is the feature representation of the k -th interest point to be recommended in the l-2th layer.

The method for obtaining the attention coefficient is as follows: in the interest interaction structure diagram of the current recommended interest point in the previous layer, the feature representation of the current interest point to be recommended in the previous layer and the feature representation of the similar users to be processed in the previous layer are sequentially processed by splicing, weighting and activation function, and then the exponential function value is obtained to obtain the attention features of the current interest point to be recommended and the similar users to be processed in the current layer; the attention features of the current interest point to be recommended in the current layer and all similar users are obtained, and after summation, the total attention features of the current interest point to be recommended in the current layer are obtained; the ratio of the attention features of the current interest point to be recommended in the current layer and the similar users to be processed to the total attention features of the current interest point to be recommended in the current layer is used as the attention coefficient of the current interest point to be recommended in the current layer and the similar users to be processed.

Specifically, the attention coefficient can be obtained by the following formula:

,

α ^ij,l is the attention coefficient between the ith interest point to be recommended and the jth similar user in the l -1th layer, m ^l-1 (u _j ) is the similar user feature representation of the jth similar user in the l-1th layer, m ^l-1 (pi ₎ is the feature representation of the interest point to be recommended of the ith interest point to be recommended in the l -1th layer, W _i,l is the weight matrix of the ith interest point to be recommended in the l-1th layer, W _j,l is the weight matrix of the jth similar user in the l- 1th layer, a is the learning parameter, Relu() is the Relu activation function; w _ij is the updated edge weight between the ith interest point to be recommended and the jth similar user, which is the product of the interaction frequency between the ith interest point to be recommended and the jth similar user, and the user personality influence of the jth similar user.

Next, the feature representation of each POI to be recommended is matched with the feature representation of the random group and processed by probability mapping to obtain the predicted score.

The probability mapping process can be achieved through the following formula:

,

yi ^is the predicted score of the i -th interest point to be recommended, Sigmoid() is the sigmoid function, g is the random group fitting feature representation, and hpi ^is the feature representation of the i -th interest point to be recommended.

Finally, the recommended points of interest corresponding to the predicted scores greater than the score threshold are taken as preferred points of interest; all the preferred points of interest form a preferred point of interest set.

S4. Based on the predicted score of each preferred point of interest in the preferred point of interest set, the expected value of each random user for each preferred point of interest is obtained; based on the expected value of each random user for each preferred point of interest, a multi-negotiation recommendation process is performed to obtain the optimal recommended point of interest for recommendation to the random group.

According to the predicted scores of the preferred points of interest, the expected value of each random user for the preferred points of interest is obtained, and based on the expected value of each random user, a multi-negotiation recommendation process is performed on the preferred points of interest to obtain the optimal recommended points of interest that satisfy all random users and are used to recommend to the random group.

The method for obtaining the expected value is: if the random user has evaluated the preferred interest point, the score of the random user on the preferred interest point is used as the expected value; if the random user has not evaluated the preferred interest point, the predicted score corresponding to the preferred interest point is used as the expected value.

After each random user obtains the expected value for each preferred point of interest, the corresponding preferred points of interest are arranged in descending order of the expected value for use in recommending points of interest in the multi-negotiation recommendation process.

The multi-negotiation process is as follows.

S4.1. In the random group, each random user selects the preferred interest point corresponding to the maximum expected value as their own recommended item; all recommended items form a recommended item set.

In the first round of negotiation of multi-negotiation processing, each random user displays the preferred interest points corresponding to their maximum expected values, that is, the priority interest points that they most want to recommend, as their respective recommended items in the random group. All recommended items form a recommended item set.

S4.2. Determine whether there is a corresponding recommended item in the recommended item set whose expected value is not less than the expected value of other recommended items for the same user; if so, take the corresponding recommended item whose expected value is not less than the expected value of other recommended items for the same user as the optimal recommended interest point for recommendation to the random group; if not, execute S4.3.

First, obtain the expected value of each random user for each recommended item in the recommended item set, and form a recommended item expectation summary table with the a-th random user as the recommended item and row a and column b as the expected value of the a-th random user for the b-th recommended item. Then, through the recommended item expectation summary table, determine whether there is a corresponding recommended item in the recommended item set with an expected value not less than the expected value of other recommended items of the same user.

If it exists, that is, there is a recommended item whose expected value in the corresponding column of the recommended item expectation summary table is greater than the expected value of other recommended items in the corresponding row (the same user), then the corresponding recommended item whose expected value is not less than the expected value of other recommended items of the same user will be regarded as the recommended item accepted by all random users, that is, as the optimal recommended interest point, for recommendation to the random group.

If not, it is necessary to execute the next round of multi-negotiation processing, that is, execute S4.3.

S4.3. Based on the recommended item set and the corresponding expected value, obtain the risk-taking willingness of each random user; take the random user corresponding to the minimum risk-taking willingness as the target user; select a preferred interest point corresponding to a target user whose expected value is not the maximum value as the target recommended item; replace the recommended item of the target user in the recommended item set with the target recommended item to obtain the negotiated recommended item set for executing the operation of S4.2.

If there is no recommended item whose expected value in the corresponding column of the recommended item expectation summary table is not less than the expected value of other recommended items in the corresponding row, then firstly obtain the risk-taking willingness of each random user based on the recommended item set and the corresponding expected value.

The method for obtaining the risk-taking willingness is as follows: if the expected value of the random user for the recommended item set is 0, the risk-taking willingness of the random user is 1; if the expected value of the random user for the recommended item set is not 0, the ratio of the difference between the expected average value of the random user and the minimum expected value in the recommended item set to the expected average value of the random user is taken as the risk-taking willingness of the random user.

Then, a random user corresponding to the minimum risk-taking willingness is taken as the target user, and a preferred interest point corresponding to a target user whose expected value is not the maximum is selected as the target recommendation item. The preferred interest point with the second expected value can be selected as the target recommendation item in descending order of expected values.

Finally, the target recommended item replaces the recommended item corresponding to the target user in the recommended item set in the previous round of negotiation processing, and the negotiated recommended item set is obtained for executing the operation of S4.2. That is, the second round of multi-negotiation processing is executed, and the operation of judging the expected value of the recommended item is executed. If there is still a recommended item whose expected value in the corresponding column of the recommended item expectation summary table is not less than the expected value of other recommended items in the corresponding row, then the risk willingness calculation, target user selection and update of the recommended item set in S4.3 are executed.

In addition, in order to make the multi-negotiation process closer to the actual situation and improve the accuracy of the recommendation. In S4.3, if the user personality influence of the target user is less than the influence threshold, the target user is more likely to make concessions during the negotiation process. In order to consider the feelings of the target user, the expected values of other similar users for the target recommendation items are updated respectively to obtain an updated expected value; the updated expected value is the sum of the initial expected value and the expected value compensation.

To verify the effect of the method recommended in this embodiment, the following experiment was conducted.

Datasets and experimental environment. The three datasets used in the experiment are the Yelp dataset, the Gowalla dataset, and the Foursquare dataset. These three datasets contain a large amount of user check-in data, and each check-in data contains a unique user ID and the ID of the point of interest to be recommended. The experiment performed data preprocessing operations on the three datasets to filter out inactive users and unpopular points of interest, and randomly generated the Big Five personality information for each user during the preprocessing process. Finally, 60% of the dataset was selected as the training set, 20% of the dataset was used as the validation set, and the remaining 20% of the dataset was used as the test set. Table 1 gives the detailed information of the three datasets after preprocessing.

Table 1 Statistics of Foursquare, Gowalla, and Yelp datasets

.

The hardware environment of the experiment is as follows: operating system: Windows 10 Professional Edition 64-bit; CPU: 12th GenIntel(R) Core(TM) i9-12900K 3.19 GHz; GPU: NVIDIA GeForce RTX 3090; RAM: 32GB; Pycharm (Community Edition) was selected as the development platform, and Python 3.8 was used to implement the model based on the Pytorch deep learning framework.

Experiment 1: Evaluation of the effect of POI recommendation method

In order to evaluate the POI recommendation performance of the method in this embodiment, the experiment selected Precision@K and Normalized Discounted Cumulative Gain (NDCG@K) as indicators for evaluation. The calculation formulas of the two indicators are as follows:

,

,

,

R(g) is the recommended list of points of interest, T(g) is the real check-in list of points of interest of the random group RG , rel _is a binary value used to indicate whether the th point of interest in the recommended list is in the real list, DCG@K is the discounted cumulative gain, IDCG@K is the ideal discounted cumulative gain, which represents the maximized DCG@K value, and the larger the NDCG@K value, the better the recommendation method.

Existing methods for comparison. Existing group POI recommendation mainly conducts POI recommendation research for fixed groups, and POI recommendation research for random groups is relatively rare. Therefore, it is difficult to find a more ideal existing method to verify the recommendation performance of the method of this embodiment. Considering that the method of this embodiment obtains random group fitting features by feature aggregation, and then recommends POIs based on the group fitting features, this processing flow is essentially similar to personalized POI recommendation. Therefore, 8 personalized POI recommendation models were selected in the experiment as existing methods for performance comparison with the method of this embodiment. These 8 existing methods are as follows.

TransMKR: TransMKR is a multi-task learning model for POI recommendation based on knowledge graph translation. The model uses the attributes of POI (number of visits to POI, geographic location of POI, etc.) to construct a knowledge graph. Different attribute values can characterize the characteristics of POI in a more detailed and accurate manner.

NPGR: NPGR constructs a heterogeneous LBSN graph that includes users, categories of POIs, and check-in time windows. The Node2Vec method is used to extract node feature representations. In addition, NPGR also takes into account factors such as user check-in frequency, geographic location of POIs, and popularity of POIs into consideration for POI recommendation.

GSTN: GSTN is a graph-enhanced spatiotemporal network model. With the help of graph neural network technologies, GSD can effectively capture the correlation between time and space in points of interest.

STORE: STORE studies the impact of time and space factors on user check-in behavior in interesting point recommendation. Compared with the traditional method of studying time and space separately, STORE studies time and space together.

MBR: MBR is a multivariate bipartite graph neural network POI recommendation model that effectively reduces the computational overhead of traditional bipartite graph models by performing clustering operations on the graph model. In order to improve the performance of the recommendation model, MBR deeply analyzes the impact of the user's social network, the geographic location of the POI, and the user's check-in time on the user's check-in behavior.

STGN: The STGN model applies the gating mechanism to the study of spatiotemporal factors. Considering that the user's check-in behavior is a serialized behavior, and the gating mechanism is one of the most efficient tools for processing serialized data, the introduction of the gating mechanism to model the user's check-in behavior pattern effectively improves the accuracy of the recommendation.

FG-CF: FG-CF combines collaborative filtering and graph convolutional networks to effectively alleviate the problem of sparse user check-in data in POI recommendation. At the same time, the user's social information is added to the user-POI bipartite graph, which can more accurately characterize the user's characteristics.

DSMR: DSMR improves the quality of trajectory embedding by learning deep semantic information from discrete trajectory data. Specifically, it performs continuous semantic modeling on discrete trajectory data and uses a pre-trained language model to extract its implicit deep semantic information to further improve the recommendation performance.

TGSTAN: TGSTAN considers collaborative signals and dynamic user preferences in graph learning, and proposes a GCN model combined with self-attention to improve the model recommendation performance by emphasizing the relative proximity of spatiotemporal intervals.

Performance comparison of recommendation methods. In order to verify the performance of the point of interest recommendation of the method of this embodiment, the point of interest recommendation effect of the method of this embodiment on the three data sets of Foursquare data set, Gowalla data set and Yelp data set was compared with the existing methods. In particular, since not all existing methods are experimented on the three data sets of Foursquare data set, Gowalla data set and Yelp data set, and the normalized cumulative loss gain values of some existing methods are not consistent with the normalized cumulative loss gain values of the method of this embodiment. Therefore, the experimental idea of performance comparison in the experiment is to compare the method of this embodiment with the existing methods using the same data set and the same normalized cumulative loss gain value. The final comparison results are shown in Tables 2, 3 and 4. Among them, "-" indicates that the model lacks the recommendation results of the corresponding indicators. On the Foursquare data set, FG-CF, GSTN, NPGR, TransMKR, DSMR and TGSTAN were selected as comparison methods for existing methods, and performance comparison experiments were carried out with the method of this embodiment. It can be seen from Table 2 that for the evaluation index accuracy, when K is equal to 2, the accuracy reaches a maximum value of 0.97; when K is equal to 20, the accuracy value is reduced to 0.85. For the evaluation index normalized cumulative loss gain, its value reaches a maximum value of 0.96 when K is equal to 2, and then as K increases, the value of the normalized cumulative loss gain gradually decreases. When K is equal to 20, the normalized cumulative loss gain reaches a minimum value of 0.87. Based on the performance comparison results in Table 2, it can be calculated that compared with the existing method, the method of this embodiment has improved the accuracy and normalized cumulative loss gain by 50% and 62% respectively. The recommendation performance of the method of this embodiment is better than the existing method as a whole.

Table 2 Performance comparison results of the method in this embodiment and the existing method on the Foursquare dataset

.

On the Gowalla dataset, this paper selected FG-CF, GSTN, MBR, TransMKR and TGSTAN as comparison methods for existing methods, and conducted performance comparison experiments with the method of this embodiment. It can be seen from Table 3 that when K increases from 2 to 20, the value of the accuracy shows a trend of fluctuating up and down. Specifically, when K increases from 2 to 5, the accuracy decreases from 0.88 to 0.80; then, when K increases from 5 to 15, the accuracy gradually increases from 0.80 to 0.86; finally, when K increases from 15 to 20, the accuracy gradually increases from 0.86 to 0.85. For the evaluation index normalized discounted cumulative gain, when K increases from 2 to 20, the value of the normalized discounted cumulative gain shows an overall trend of first decreasing and then increasing. Specifically, when K increases from 2 to 5, the normalized cumulative loss gain decreases from 0.87 to 0.82; when K increases from 5 to 20, the normalized cumulative loss gain increases from 0.82 to 0.85. Based on the performance comparison results in Table 3, it can be calculated that compared with the existing methods, the method of this embodiment improves the accuracy and normalized cumulative loss gain by an average of 61.5% and 64% respectively.

Table 3 Performance comparison results of the method in this embodiment and the existing method on the Gowalla dataset

.

In the research field of POI recommendation, the frequency of use of the Yelp dataset is lower than that of the Foursquare dataset and the Gowalla dataset. Among the existing comparison methods selected in the experiment, only the FG-CF model has been tested on the Yelp dataset. Therefore, on the Yelp dataset, this paper selected FG-CF as the existing comparison method and conducted a performance comparison experiment with the method of this embodiment. It can be concluded from Table 4 that for the evaluation index accuracy, when K increases from 2 to 20, the value of the normalized discounted cumulative gain decreases from 0.08 to 0.06. The method of this embodiment is better than the existing comparison method when the K value is 10, 15, and 20.

Table 4 Comparison results of the method in this embodiment and the existing method on the Yelp dataset

.

From the performance comparison results of Table 2, Table 3 and Table 4, it can be seen that for the random group POI recommendation problem, the method of this embodiment can achieve better recommendation effect.

Experiment 2: Evaluation of the multi-negotiation treatment experiment.

Performance evaluation indicators of multi-negotiation processing methods The multi-negotiation processing method aims to simplify the selection process of interest points in the interest point recommendation list by group members. In order to verify the effectiveness of the multi-negotiation processing method, the multi-negotiation processing method in OPRM is experimentally verified by user satisfaction, using two evaluation indicators: group satisfaction (GS) and member satisfaction dispersion (MSD). Among them, group satisfaction (GS) aims to measure the overall satisfaction of the group with the recommended interest points. Among them, satisfaction is equivalent to the estimated preference (or estimated rating) of the user or group for the project. Member satisfaction dispersion (MSD) aims to evaluate the satisfaction of group members with a single proposed interest point. The lower the MSD, the higher the satisfaction of the group members.

The evaluation index formula is as follows:

,

,

GS(p _j ) is the group satisfaction of the random group for the interest point p _j , E _i (p _j ) represents the approximate estimated preference _of user ui for the interest point p _j , MSD(p _j ) is the dispersion of the member satisfaction of the random group for the interest point p _j , and n is the current random group size.

The impact of group size on negotiation effect. In the evaluation of the effect of the multi-negotiation processing method, the Foursquare dataset is used as the data set, and the group size is set to [10, 20, 30, 40, 50, 60, 70, 80, 90, 100]. The multi-negotiation processing method of this embodiment will negotiate and deduce the points of interest in the TOP-5 recommendation sequence under different group sizes, and verify the effectiveness of the multi-negotiation processing method based on the group satisfaction of group users on the negotiated points of interest and the dispersion of member satisfaction. The experimental results are shown in Figures 2 and 3.

From Figures 2 and 3, it can be observed that on the Foursquare dataset, when the group size is in the range of [10, 100], the group satisfaction and member satisfaction dispersion show a fluctuating trend, although the overall fluctuation is small. Since the Foursquare dataset does not record the specific ratings of users for points of interest, but only records the number of times users check in to points of interest. In the experiment, when obtaining user ratings, the check-in frequency is used as the user's estimated rating. It can be noted that the value of the member satisfaction dispersion is in the range of [0.005, 0.03]. The method of this embodiment performs well in the negotiation deduction process, which means that even if the user's actual rating is low, the method of this embodiment can still achieve a good member satisfaction dispersion in group negotiation, demonstrating the superiority of the model in the negotiation process.

Effectiveness evaluation of multi-negotiation processing method. The existing group interest point recommendation method based on negotiation processing first adopts a personalized recommendation method to obtain the user's estimated preference to complete the user's rating of the interest point, and then generates a recommendation list of interest points through a negotiation method. In contrast, this embodiment adopts a different method, first obtaining a recommendation list of interest points through a group interest point recommendation method, and then obtaining a more accurate interest point recommendation through a multi-negotiation processing method. In addition, the existing group interest point recommendation method based on negotiation processing has a defect, that is, it lacks the fitting of group characteristics and cannot obtain the overall preference of the group, thereby reducing the recommendation effect. In addition, the precise interest point recommendation method of first recommendation and then negotiation proposed in this embodiment is proposed for the first time, so this embodiment faces the challenge of finding an ideal existing method to verify the synergistic effect of multi-negotiation. In order to determine the effectiveness of the multi-negotiation processing method, the group satisfaction of randomly selected interest points and interest points derived through multi-negotiation processing is used as an evaluation indicator in the experiment, and the experimental results are shown in Figure 4. As can be seen from Figure 4, the interest points obtained by the multi-negotiation processing method of this embodiment are greater than or equal to the randomly selected interest points in terms of group satisfaction. It has been proved that compared with randomly selecting the final signed-in points in the POI recommendation sequence, the multi-negotiation processing method of this embodiment can not only achieve accurate selection of POIs in the POI recommendation sequence, but also improve the user experience of group members.

This embodiment also provides a random group point of interest recommendation system, including:

The module for generating the set of interest points to be recommended is used to obtain similar users of each random user in the random group, and all similar users form a similar user set; the product of the user feature vector of the similar user and the feature vector of the random user is greater than the similarity threshold; the interest points that each similar user has interacted with in the similar user set are obtained as the interest points to be recommended; all the interest points to be recommended form a set of interest points to be recommended;

The random group fitting feature representation generation module is used to obtain the relative influence weight of each random user in the random group based on the set of interest points to be recommended; based on the relative influence weight of each random user and the influence of each user's personality, the absolute influence weight of each random user is obtained; the absolute influence weight of each random user is multiplied by the respective random user feature vector, and then aggregated to obtain the random group fitting feature representation;

The preferred interest point set generation module is used to construct an interest interaction structure diagram for each recommended interest point based on the interest points to be recommended; the interaction structure diagram is composed of the interest points to be recommended and/or similar users as nodes, and the interaction edges between the interest points to be recommended and similar users; the interest interaction structure diagram of each recommended interest point is processed by a multi-layer graph neural network to obtain the feature representation of each interest point to be recommended; the feature representations of all the interest points to be recommended form a feature representation set of the interest points to be recommended; in the feature representation set of the interest points to be recommended, the feature representation of each interest point to be recommended is respectively matched with the feature representation of the random group fitting, and a predicted score is obtained by probability mapping; the interest points to be recommended corresponding to the predicted score greater than the score threshold are taken as the preferred interest points; all the preferred interest points form a preferred interest point set;

The multi-negotiation recommendation processing and interest point recommendation module is used to obtain the expected value of each random user for each preferred interest point based on the predicted score of each preferred interest point in the preferred interest point set; based on the expected value of each random user for each preferred interest point, multi-negotiation recommendation processing is performed to obtain the optimal recommended interest point for recommendation to the random group;

The multi-negotiation recommendation processing and POI recommendation modules include:

The recommended item set generation module is used in a random group. Each random user selects the preferred interest point corresponding to the maximum expected value as their own recommended item. All recommended items form a recommended item set.

The expected value judgment module is used to judge whether there is a corresponding recommended item in the recommended item set whose expected value is not less than the expected value of other recommended items for the same user; if so, the corresponding recommended item whose expected value is not less than the expected value of other recommended items for the same user is used as the optimal recommended interest point for recommendation to the random group; if not, the negotiation recommended item set generation module is executed;

The negotiated recommended item set generation module is used to obtain the risk-taking willingness of each random user based on the recommended item set and the corresponding expected value; the random user corresponding to the minimum risk-taking willingness is taken as the target user; a preferred interest point corresponding to a target user whose expected value is not the maximum is selected as the target recommended item; the target recommended item replaces the recommended item of the target user in the recommended item set with the target recommended item to obtain the negotiated recommended item set, which is used to execute the operation of the expected value judgment module.

This embodiment further provides a device for recommending points of interest for a random group, including a processor and a memory, wherein the processor implements the above-mentioned method for recommending points of interest for a random group when executing a computer program stored in the memory.

This embodiment further provides a computer-readable storage medium for storing a computer program, wherein the computer program implements the above-mentioned random group point of interest recommendation method when executed by a processor.

The present embodiment provides a method for recommending points of interest for a random group. First, based on the size relationship between the product of feature vectors and a similarity threshold, similar users with similar preferences as the random users in the random group are selected, and the points of interest that the similar users have interacted with are obtained as the recommended points of interest that the random group may be interested in, thereby narrowing the scope of selecting points of interest. Then, according to the characteristics of the set of points of interest to be recommended, the relative influence weight of each random user is obtained, and the relative influence weight of each random user is combined with the user personality influence that can reflect the respective personality characteristics to obtain the absolute influence weight of each random user in the random group. After multiplying the absolute influence weight of each random user with the respective random user feature vector, aggregation processing is performed to obtain a random group fitting feature representation that can reflect the overall preference of the random group. Next, each point of interest to be recommended and the similar users with whom it has interacted are constructed into a structural graph, and a multi-layer graph is performed. Neural network processing is used to realize the message propagation between the recommended interest points and similar users, and a feature representation of each interest point to be recommended with rich features and stronger expression ability is obtained; then, each feature representation of the interest point to be recommended is probability mapped with the random group fitting feature representation to obtain a predicted score; and the recommended interest points with predicted scores greater than the score threshold are used as preferred interest points that are more easily accepted by the random group, further narrowing the range of selected interest points; finally, according to the predicted scores of the preferred interest points, the expected values of each random user for the preferred interest points are obtained, and based on the expected values of each random user, multi-negotiation recommendation processing is performed on the preferred interest points to obtain the optimal recommended interest points that satisfy all random users for recommendation to the random group; this method screens and negotiates interest points based on the user's personality, and the interest point prediction results are more accurate, the recommendation accuracy is higher, and the user's commercial experience can be improved.

Claims (8)

1. A method for recommending points of interest of a random group, characterized by comprising the following operations: S1. Obtain similar users of each random user in the random group, and all similar users form a similar user set; the product of the user feature vector of the similar user and the feature vector of the random user is greater than the similarity threshold; obtain the points of interest that each similar user has interacted with in the similar user set as the points of interest to be recommended; all the points of interest to be recommended form a set of points of interest to be recommended; S2. Based on the set of interest points to be recommended, the relative influence weight of each random user in the random group is obtained; the relative influence weight of the random user is obtained by: obtaining the differential vectors of the random user to be processed and other random users in the random group, and fusing all the differential vectors to obtain the fused vector to be processed; obtaining the fused vector to be processed, and multiplying it by the feature vector of the current interest point to be recommended in the set of interest points to be recommended, to obtain the current comparison vector to be processed; all the comparison vectors to be processed are sequentially processed by splicing, nonlinear processing and activation function, to obtain the relative influence weight of the random user to be processed; Based on the relative influence weight of each random user and the influence of each user's personality, the absolute influence weight of each random user is obtained; the method for obtaining the absolute influence weight of the random user is as follows: a random user is randomly selected from the random group as a representative random user; the absolute value of the difference between the relative influence weight of other random users and the relative influence weight of the representative random user is used as the absolute factors of other random users; the absolute factors of other random users are multiplied by their respective user personality influences to obtain the absolute influence weights of other random users; the relative influence weight of the representative random user is multiplied by the corresponding user personality influence to obtain the absolute influence weight of the representative random user; The absolute influence weight of each random user is multiplied by the respective random user feature vector and then aggregated to obtain the random group fitting feature representation; S3. Based on the interest points to be recommended, construct an interest interaction structure graph of each recommended interest point; the interaction structure graph is composed of the interest points to be recommended and/or similar users as nodes, and the interaction edges between the interest points to be recommended and the similar users; The interest interaction structure graph of each recommended interest point is processed by a multi-layer graph neural network to obtain the feature representation of each interest point to be recommended; the feature representations of all interest points to be recommended form a feature representation set of interest points to be recommended; In the feature representation set of the interest points to be recommended, each feature representation of the interest points to be recommended is respectively matched with the feature representation of the random group fitting, and a predicted score is obtained through probability mapping processing; the interest points to be recommended corresponding to the predicted score greater than the score threshold are taken as the preferred interest points; all the preferred interest points form a preferred interest point set; S4. Based on the predicted score of each preferred point of interest in the preferred point of interest set, obtain the expected value of each random user for each preferred point of interest; based on the expected value of each random user for each preferred point of interest, perform multi-negotiation recommendation processing to obtain the optimal recommended point of interest for recommendation to the random group. 2. The method for recommending points of interest of a random group according to claim 1, wherein the operation of the multi-negotiation recommendation process in S4 is specifically: S4.1. In the random group, each random user selects the preferred interest point corresponding to the maximum expected value as the recommended item; all the recommended items form a recommended item set; S4.2, determining whether there is a corresponding recommended item in the recommended item set whose expected value is not less than the expected value of other recommended items for the same user; if so, taking the corresponding recommended item whose expected value is not less than the expected value of other recommended items for the same user as the optimal recommended interest point for recommendation to the random group; if not, executing S4.3; S4.3. Based on the recommended item set and the corresponding expected value, the risk-taking willingness of each random user is obtained; the random user corresponding to the minimum risk-taking willingness value is taken as the target user; a preferred interest point corresponding to the target user's expected value which is not the maximum value is selected as the target recommended item; the target recommended item replaces the recommended item of the target user in the recommended item set to obtain a negotiated recommended item set for executing the operation of S4.2. 3. The method for recommending points of interest of a random group according to claim 1, wherein the method for obtaining the influence of the user's personality is: The user's score under the Big Five personality model is multiplied by the corresponding personality influence ratio, and then summed and normalized to obtain the user's personality influence. 4. The method for recommending points of interest of a random group according to claim 1, characterized in that, in S4, the method for obtaining the expected value is: If the random user has evaluated the preferred interest point, the random user's rating of the preferred interest point is used as the expected value; If the random user has not evaluated the preferred interest point, the predicted score corresponding to the preferred interest point is used as the expected value. 5. The method for recommending points of interest for a random group according to claim 1, characterized in that, in S4, the method for obtaining the risk-taking willingness is: If the expected value of a random user for the recommended item set is 0, then the random user's willingness to take risk is 1; If the expected value of the random user for the recommended item set is not 0, the ratio of the difference between the expected average value of the random user and the minimum expected value in the recommended item set to the expected average value of the random user is used as the risk-taking willingness of the random user. 6. A random group point of interest recommendation system, characterized by comprising: The module for generating a set of interest points to be recommended is used to obtain similar users of each random user in the random group, and all similar users form a similar user set; the product of the user feature vector of the similar user and the feature vector of the random user is greater than the similarity threshold; the interest points that each similar user has interacted with in the similar user set are obtained as the interest points to be recommended; all the interest points to be recommended form a set of interest points to be recommended; The random group fitting feature representation generation module is used to obtain the relative influence weight of each random user in the random group based on the set of interest points to be recommended; the method for obtaining the relative influence weight of the random user is as follows: the differential vectors of the random user to be processed and other random users in the random group are obtained, and all the differential vectors are fused to obtain the fused vector to be processed; the fused vector to be processed is obtained, and the product of the feature vector of the current interest point to be recommended in the set of interest points to be recommended is obtained to obtain the current comparison vector to be processed; all the comparison vectors to be processed are sequentially processed by splicing, nonlinear processing and activation function processing to obtain the relative influence weight of the random user to be processed; based on the relative influence weight of each random user and the influence of each user's personality, the relative influence weight of the random user to be processed is obtained. The absolute influence weight of each random user; the method for obtaining the absolute influence weight of a random user is as follows: a random user is randomly selected from the random group as a representative random user; the absolute value of the difference between the relative influence weights of other random users and the relative influence weight of the representative random user is used as the absolute factors of other random users; the absolute factors of other random users are multiplied by the influence of their respective user characteristics to obtain the absolute influence weights of other random users; the relative influence weight of the representative random user is multiplied by the corresponding user personality influence to obtain the absolute influence weight of the representative random user; the absolute influence weight of each random user is multiplied by the respective random user feature vector, and then aggregated to obtain a random group fitting feature representation; The preferred interest point set generation module is used to construct an interest interaction structure diagram for each recommended interest point based on the interest points to be recommended; the interaction structure diagram is composed of the interest points to be recommended and/or similar users as nodes, and the interaction edges between the interest points to be recommended and similar users; the interest interaction structure diagram of each recommended interest point is processed by a multi-layer graph neural network to obtain a feature representation of each interest point to be recommended; all the feature representations of the interest points to be recommended form a feature representation set of the interest points to be recommended; in the feature representation set of the interest points to be recommended, each feature representation of the interest point to be recommended is respectively matched with the random group fitting feature representation, and a predicted score is obtained by probability mapping; the interest points to be recommended corresponding to the predicted score greater than the score threshold are taken as preferred interest points; all the preferred interest points form a preferred interest point set; The multi-negotiation recommendation processing and interest point recommendation module is used to obtain the expected value of each random user for each preferred interest point based on the predicted score of each preferred interest point in the preferred interest point set; based on the expected value of each random user for each preferred interest point, multi-negotiation recommendation processing is performed to obtain the optimal recommended interest point for recommendation to the random group. 7. A device for recommending points of interest for a random group, comprising a processor and a memory, wherein the processor implements the method for recommending points of interest for a random group as described in any one of claims 1 to 5 when executing a computer program stored in the memory. 8. A computer-readable storage medium, characterized in that it is used to store a computer program, wherein when the computer program is executed by a processor, the method for recommending points of interest for a random group according to any one of claims 1 to 5 is implemented.