CN108492200B - User attribute inference method and device based on convolutional neural network - Google Patents

Abstract

The invention relates to a method and device for inferring user attributes based on a convolutional neural network. The method establishes a self-centered network according to the attributes and friend relationships of user nodes; then uses a convolutional neural network to extract the attribute information of the user nodes in the self-centered network and the hidden information contained in the friend relationships, and uses the hidden information to infer out the missing attributes of the user. For social networks where friend relationships cannot be directly acquired or are difficult to acquire, neural networks are used to classify and predict missing attributes using only the user's attribute information. The present invention can well avoid the limitation of artificially defining the similarity function, and can better represent the relationship between different attributes and between different attribute dimensions through the convolution operation of the convolution kernel, so that it can be efficiently and accurately performed. User-missing attribute inference.

Description

A method and device for user attribute inference based on convolutional neural network

technical field

The invention belongs to the technical field of user missing attribute inference in social networks, in particular to the inference of multi-valued attributes, and in particular relates to a method and device for inferring user attributes based on a convolutional neural network, which has high accuracy.

Background technique

With the development of Internet technology, online social products, such as Weibo, Zhihu and Facebook, have become the necessities of users' daily life. While using these products, users generate a large amount of information, including the user's attribute information, post content and friend relationships, which provide data support for enterprises and researchers to accurately describe user portraits. At the same time, in order to protect the privacy of users, online social products provide users with fine-grained privacy settings, which makes it difficult to obtain user attribute information directly. According to relevant statistics, the missing rate of user attribute information is as high as 70%. The lack of a large amount of information has become the biggest problem in accurately describing user portraits. Therefore, user attribute inference technology has attracted extensive attention in industry and scientific research.

Traditional methods can be divided into two categories - classification-based methods and label propagation-based methods. The theoretical basis of the former is what you belong to, usually predicting unknown properties by computing the similarity of a marked node to its neighbors. The calculation of similarity mainly depends on the specific method used. The classic classification algorithms include SVM, Bayesian and so on. In addition, some scholars have proposed a similarity calculation method that is more in line with the actual situation. A better method for calculating similarity, with good results. However, in general, models such as SVM and Bayesian have good performance in predicting attributes such as gender and age, but their performance in predicting occupation, interests and other issues is not satisfactory. The method based on tag propagation takes advantage of the homogeneity of social networks, that is, two users with friend relationships are more likely to have similar attributes. Based on this theoretical basis, attribute values can be propagated from nodes with known attribute information to nodes with unknown attribute information through edges, so as to accurately predict attribute values. For example, in the article "Exploit of Online Social Networks with Semi-Supervised Learning" proposed by Mingzhen Mo in 2010, the method of mark propagation is used to predict the unknown attributes of users, and good results have been achieved. However, in practice, this method requires a lot of time and space overhead to calculate the adjacency matrix of the graph composed of social networks.

According to the above introduction, there are many researches on user missing attribute filling at home and abroad, which are mainly divided into classification-based and tag-based propagation. The limitation of the classification-based method is that the calculation of similarity needs to be comprehensive and accurate, the classification model must have high accuracy, and the construction of data features must be comprehensive. However, although the accuracy of the machine learning model in the current method is high, most The problem of binary classification, and the inability to obtain multi-dimensional attributes to construct a comprehensive feature vector of users in social networks, etc., leads to poor prediction effects, especially the prediction of multi-valued attributes. The attribute prediction algorithm based on mark propagation needs to spend a lot of time to calculate the adjacency matrix of the graph, and the algorithm itself treats the importance of the friends of the marked node equally, which itself is not in line with the characteristics of social networks, so on the real data The effect is also unsatisfactory.

In addition, when predicting unknown attributes of marked nodes, these two types of methods need to train models according to specific attributes. For example, when predicting gender attributes, a corresponding model needs to be established according to the user's attribute information. When predicting occupational attributes, The previously trained model cannot be used directly, and a new model needs to be trained to achieve a better prediction effect.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is to provide an efficient and accurate user missing attribute inference technology, so as to better describe the user portrait.

The technical scheme adopted in the present invention is as follows:

A method for user attribute inference based on convolutional neural network, including the following steps:

1) According to the attributes and friend relationships of user nodes, establish a self-centered network;

2) Using a convolutional neural network to extract the attribute information of the user node in the self-center network and the hidden information contained in the friend relationship, and use the hidden information to infer the missing attribute of the user.

Further, the self-central network is represented by a quintuple G'={V',',',',}, where V' includes node information in the self-central network, and the set E' includes all nodes in the self-central network The link relationship between the nodes, the sets A' and ' respectively represent the attribute information and behavior information of the node, and the matrix L∈V'×N contains the attribute information and behavior information of the self-center node and its friends, which is the sum of the attributes and behavior data dimensions. .

Further, step 1) firstly filters the attribute information of the user on the network, and then establishes the self-centered network; the filtering includes:

a) Filter out all non-Chinese character phrases with attributes other than age.

b) Filter out nodes whose attribute information missing exceeds the set threshold.

Further, the convolutional neural network includes an input layer, a projection layer, a convolution layer, a pooling layer, a fully connected layer and an output layer, and the projection layer converts the user's attribute information and behavior information into vectors.

Further, the projection layer converts the user's attribute information and behavior information into vectors, and adopts the following rules for words belonging to the same occupation or profession:

i. Create a hash table, take the first word of the profession or occupation as the key, and use the set of all attributes with the first word as the value as the value; then calculate the similarity of other attribute values with the words in the professional and occupation dictionary according to the jara-winkler distance properties, and add attributes with high similarity to the set value;

ii. For those attribute values that do not appear in the set value, use word2vec to train word vectors. By calculating the distance between the word vectors, the vectors with high similarity are aggregated together by the KNN algorithm, and the words in the dictionary are passed ID numbers. Correlate to get a digitized vector.

Further, the convolutional layer uses the trained weight matrix and bias to map with the feature matrix obtained after the projection layer, and uses Relu as the activation function; the pooling layer uses maximum pooling to retain local features. The most important information in the output layer; the output layer uses the softmax classifier to score the possible values of the attribute by using the weight matrix and the bias value, and the attribute value with the highest score is the missing attribute value of the marked user.

Further, for social networks in which the friend relationship cannot be directly obtained or is difficult to obtain, the neural network is used to classify and predict the missing attributes only by using the user's attribute information.

Further, the neural network includes an input layer, a projection layer, a hidden layer and an output layer; the projection layer converts the user's attribute information and behavior information into vectors; the hidden layer is two fully connected layers, the first one The hidden layer contains n*n neurons, where n is the sum of the dimensions of the user's attributes and behavior data, and the second hidden layer loses some neurons to prevent overfitting; the output layer uses the softmax classifier to use the weight matrix and The biased value is used to score the possible values of the attribute, and the attribute value with the highest score is the missing attribute value of the marked user.

A device for inferring user attributes based on a convolutional neural network, comprising:

The self-centered network building module is responsible for establishing a self-centered network according to the attributes and friend relationships of user nodes;

The user attribute inference module is responsible for extracting the attribute information of the user node in the self-center network and the hidden information contained in the friend relationship by using a convolutional neural network, and using the hidden information to infer the missing attribute of the user.

Further, for social networks in which the friend relationship cannot be directly obtained or is difficult to obtain, the user attribute inference module uses a neural network to classify and predict the missing attributes only by using the user's attribute information.

The user attribute inference algorithm based on the convolutional neural network proposed by the present invention is similar to the classification-based method in the prior art, and the attribute inference problem is also regarded as a classification problem. The biggest difference is that the traditional classification method will define a similarity calculation formula. According to the formula, the similarity between nodes is calculated and the possible values of the unknown attributes are scored, so that the attribute value with the highest score is selected as the Unknown attribute value of the marker node. However, the attribute prediction algorithm based on the convolution network of the present invention, UPE, calculates the score of the possible value of the unknown attribute of the marked node according to the weight matrix and the offset value. Take the potential relationship between different attribute dimensions, and then calculate the weight matrix and bias of the attribute according to these relationships and the back-propagation algorithm. After a linear calculation, the score of each possible value of the unknown attribute is obtained. The method of measuring similarity by weight matrix and bias can well avoid the limitation of artificially defined similarity function, and the convolution operation of convolution kernel can better show the difference between different attributes and between different attribute dimensions. This relationship cannot be explained and measured, and it is also ignored by traditional classification algorithms. In addition, this is the first attempt of convolutional neural network on the problem of attribute inference since it was proposed, and has achieved good experimental results.

In addition, as a supplement to the exception of the UPE algorithm, the neural network approach adopted by UPS is also applied to the problem of attribute inference for the first time. The biggest difference from the prior art method is that the neural network method is more flexible and efficient, and there is no fixed number of neural network layers and similarity functions, so when dealing with complex and changeable data problems in some social networks It shows good inclusiveness, and the effect is better than other classification algorithms. Through testing and tuning on real datasets, a UPS structure diagram is constructed.

Description of drawings

Figure 1: User-wide network diagram and self-centered network diagram.

Figure 2: UPS flow chart.

Figure 3: UPE flow chart.

Figure 4: A comparison diagram of the experimental results of the present invention and other machine learning algorithms.

Detailed ways

In order to make the above objects, features and advantages of the present invention more clearly understood, the present invention will be further described in detail below through specific embodiments and accompanying drawings.

The convolutional neural network-based user missing attribute inference technology of the present invention includes the following steps:

Step 1, dimension filtering:

According to the user's attribute information on the network, the analysis shows that the user's attributes mainly have 15 dimensions, including user name, gender, age, school, major, profile, occupation, location, etc. In order to better implement the algorithm, the attribute information is first filtered. The specific methods are as follows:

1) Filter out all non-Chinese character phrases with attributes other than age. Age is an important dimension of user attributes, but the information description of other attributes except age is composed of Chinese characters. If it is composed of non-Chinese characters, it will not only be unable to obtain useful information from it, but it will also have a bad effect on the experiment. Therefore, filtering out this information is to better implement the attribute prediction algorithm.

2) Filter out nodes whose attribute information missing exceeds the set threshold. For example, the threshold is set to 50%. Although this algorithm is designed to solve the problem of missing attributes, for those nodes whose attribute information is less than 50%, if there is no other information such as the content of the relevant postings, it is impossible to pass the friend relationship and other information only. Known attribute information to make effective predictions.

Step 2, dictionary building and social network definition:

Based on the professional discipline settings of colleges and universities in my country, the national occupation standard catalogue and the corresponding naming rules of Chinese abbreviations, professional and occupational dictionaries are constructed respectively, and they are associated with the dictionaries through a unique ID number (each word in the dictionary has an ID number) ; In order to improve the accuracy of the algorithm, when the identified noun is not in the above two dictionaries, the similarity between the word and the word in the dictionary is calculated in turn according to the jaro-winkler distance, and the word with the highest similarity is selected as the feature.

The jaro-winkler distance is a method to calculate the similarity between two strings. Suppose two strings S ₁ and S ₂ are given, the distance between the two strings is:

Where m is the number of matched characters, t is the number of transpositions, and |S ₁ | and |S ₂ | represent the number of characters of strings S ₁ and S ₂ respectively, that is, the length. The transposition refers to movement or exchange between characters, and the number of transpositions refers to the sum of the number of steps of movement or exchange.

In order to better describe the algorithm, two social network definition methods are proposed. One is a network composed of user attribute information, which is called a full-social-graph (full network graph). The other is a network composed of user's attribute information and friend relationship, which is called ego-network (self-centered network).

Definition 1. Full-social-graph

A complete social network can be represented by a quadruple = {V, A, B, F}. where is the set of all nodes in the social network and assumes that the number of nodes in the network is m, the set sum represents the attribute information and behavior information of the node respectively, such as A _ij and B _ij respectively represent the jth dimension attribute value of the node V _i and behavioral values. The definition contains all the information of the user in the social network, F={F ₁ ,F ₂ ,...F _i ...,F _m }, where F _i =A _i ∪B _i . The behavior information refers to a series of behaviors in social networks, such as posting content, comments, likes, and forwarding.

Definition 2.Ego-network

Ego-network, also known as the self-centered network, is composed of the attributes of the self-centered node, the friend relationship and the attribute information of its friends. In order to better describe the user's self-centered network, a five-tuple G'= {V', E', A', B', L} to represent the user's self-centered network. Among them, V' is the set of all nodes in the self-central network, and the set E' contains the link relationship between all nodes in the self-central network. The definition of sets A' and B' is similar to that of full-social-graph, that is, they represent the attribute information and behavior information of nodes respectively, while the matrix L∈V'×N contains the attribute information and behavior of the self-center node and its friends. Information, where is the sum of attributes and behavioral data dimensions.

The user's full-social-graph and ego-network are shown in Figure 1, respectively, where the network composed of nodes v ₁ , v ₂ , v ₃ and the connections between them is the user The ego-network of v1, all nodes in Figure ₁ and all attribute information related to them are the full-social-graph of users in the social network.

Step 3, user attribute inference method based on neural network:

Aiming at the problems that friend relationships in networks such as QQ and WeChat cannot be directly obtained or are difficult to obtain, the most common two-layer neural network is used to classify and predict the missing attributes only by using the user's attribute information.

Although most social networks such as Zhihu and Facebook have public friend relationships, there are still some networks such as QQ, WeChat, etc. whose friend relationships are encrypted, which makes it more difficult or even impossible to obtain friend relationship information. obtained directly. In order to solve the problem of attribute inference, the present invention designs a user attribute inference method based on neural network, called UPS. The theoretical basis of this method is the good performance of neural networks in natural language processing problems. In natural language processing, neural network algorithms usually use each sentence as a feature vector. Similarly, the present invention will use user attributes and behavior data. The formed vector is used as a feature vector, and is converted into a numeric vector by a method similar to word2vec in natural language processing, so that it can be put into the model of the neural network. Compared with traditional user-missing attribute inference methods, neural network methods can achieve higher accuracy.

The flowchart of the neural network method based on user attribute information is shown in Figure 2. The first layer of the model is to obtain the feature vector of each user, the second layer is to convert the obtained features into digital vectors, and the rest Several layers are traditional neural networks.

First, assume that any feedforward neural network with P layers can be viewed as the product of p arbitrary functions f, namely:

是神经网络的每一层的输出，

是整个神经网络的输入，

其中i∈(1,n)，n是网络中节点的个数。接下来详细的介绍图2中每一层神经网络。in,

is the output of each layer of the neural network,

is the input of the entire neural network,

where i∈(1,n), where n is the number of nodes in the network. Next, each layer of the neural network in Figure 2 is introduced in detail.

1) Convert the user vector F into a feature vector (lookup table layer, projection layer)

In natural language processing, the method of training word vectors is usually used to convert each word into a vector. Similarly, the present invention designs a projection layer to complete the function of converting the user's attribute information and behavior data into vectors, mainly to find Whether there is a corresponding word in the dictionary. However, the description of the user's occupation and other information may not be unique. For example, the programmer may fill in the program ape, program yuan, etc., and other names for students include high school students, college students, freshmen, sophomores, etc., which belong to the same occupation or professional. Has the following rules:

i. The first words belonging to the same occupation or major are basically the same, such as programmers, programmers, etc. Therefore, first create a hash table, use the first word of the professional or occupational dictionary as the keyword key, and then use the set of all attributes with this word as the first word as the value. Then, the similarity between other attribute values and words in the dictionary is calculated according to the jara-winkler distance, and the attributes with high similarity are added to the set value. The principle of Jara-winkler is that the more similar the starting characters in two strings are, the more similar they are.

ii. Words belonging to the same occupation or major have similar meanings, such as student, high school student, freshman, etc. Therefore, for those attribute values that do not appear in the set value, Google's word2vec can be used to train word vectors, and by calculating the distance (cosine distance) between word vectors, the vectors with high similarity are passed through the KNN algorithm (K nearest neighbor classification algorithm) Aggregate together and associate with the words in the dictionary through ID numbers, so as to achieve the purpose of digitizing vectors.

The processing method in the neural network is transparent, so in general, for the attribute vector _i of any node i, the representation _i of the corresponding digital vector obtained by the projection layer can be expressed by the following formula:

Among them, LTF(·) represents the above-mentioned series of operation methods.

2) Hidden layer (hidden layer)

The two hidden layers behind the projection layer shown in Figure 2 are two fully connected layers. The first hidden layer contains n*n neurons (n is the sum of the dimensions of the user's attributes and behavior data), and then a part (usually half) of the neurons will be lost in the second hidden layer, doing so The purpose is to prevent overfitting. Using ReLU as the activation function, it can be expressed by the following formula:

In addition, in order to obtain the results of the neural network training process faster, the gradient descent method is used in the process of training the neural network.

3) Output layer (softmax classifier)

In order to calculate the missing attribute value of the marked node, the output layer of the neural network adopts a softmax classifier (the missing attribute filling problem can be regarded as a classification problem). In the process of training the neural network, the weight matrix W and bias b in the network are mainly calculated by the method of backpropagation. When the weight matrix and bias are determined, the Softmax classifier uses the value of the weight matrix and bias, The possible values of the attribute are scored, and the attribute value with the highest score is the missing attribute value that marks the user.

Finally, for each input vector F _i , its output in the neural network constructed based on user attributes can be expressed by the following formula:

Step 4: User attribute inference method based on convolutional neural network:

In order to make the inference more accurate, the convolutional neural network is used to design different convolution kernels to extract the user's attribute information and the hidden information contained in the friend relationship, so as to more accurately infer the user's missing attributes.

For those social networks whose friend relationships are public, using the neural network method based on user attribute information will lose a lot of information, resulting in low accuracy of prediction results, while convolutional neural networks can capture the relationship between user attributes and link relationships. Using these potential relationships can achieve the purpose of inferring the user's missing attribute values more accurately. The theoretical basis of this algorithm is the good performance of convolutional neural networks in image recognition problems. The matrix formed by the user's self-centered network is like pictures existing in social networks, which can well show the relationship between users. By extracting these differences (features), users can be well classified (attribute inference). Similar to image recognition, the self-centered network matrix should first be converted into a digital matrix before it can be put into the convolutional neural network. The matrix processing method of the self-centered network is similar to the projection layer in the UPS method, but a more complex network is used to train data and predict unknown attribute values.

The user attribute inference method based on convolutional neural network is called UPE in the present invention. The flow chart of UPE is shown in Figure 3. The first layer is the projection layer, which is the same as the projection layer of the UPS. The subsequent convolution layer, the convolution layer, contains some convolution operations. In this layer, convolution kernels of different sizes are used to extract latent features in the input data. The following is a detailed introduction of each layer in the UPE network in order.

1) Convolution layer (convolution layer)

First of all, the input layer and projection layer in the UPE network are similar to those in the UPS, that is to say, the matrix _i corresponding to the network composed of the self-center node _i is the input of the UPS, and each element constituting _i is in After passing through the projection layer, it will become the number corresponding to the element, and the digital feature matrix converted from _i is defined as ML _i . ML _i is the original input vector of the convolutionlayer. Generally, it can be expressed as:

The convolution operation can actually be summarized as a linear operation between the elements in the weight matrix W and the feature matrix _i , but this linear operation is somewhat different from the narrow linear operation concept, which requires an activation function to Play its role, otherwise the neural network in the next layer cannot be activated. Therefore, when the feature matrix _i is transferred to the convolutional layer, there will be a trained weight matrix W and bias b to make a mapping with it. This mapping can be expressed by the following formula:

The Relu in the formula is the activation function used. As a standard mapping layer, the convolutional neural network usually extracts high-dimensional features, and how these high-dimensional features help the network to predict unknown attribute values, then look at the remaining layers in the UPS. What is the effect.

2) max pooling layer

Through the above analysis, it can be known that the size of the output vector of the first layer of convolutional layer depends on the number of friends from the center node (the number of columns of the matrix), but which friend contributes the most to the unknown attributes of the self-center node, which requires The local features captured by the convolutional neural network (the size of the convolution kernel determines the range of feature extraction) are combined with the global features and then judged. The traditional convolutional neural network usually adopts average pooling or max pooling after the convolutional layer, but the average pooling is not very suitable in this case, because in general, the attributes of most friends are used for unknown attribute prediction. The value does not have much effect, so the most important information in local features can be preserved by using max pooling, which is expressed by the formula:

是卷积层p-1的输出，这一层产生出的全局特征最终会被输出到接下来的卷积层或者全连接层。UPS方法中剩下的全连接层和输出层是与UPE方法的最后两层是类似的，在这里就不在重复的介绍了。在UPS方法中，最终会通过softmax函数对属性每个可能的取值进行打分，分数高的就是要预测的未知的值，这一点是神经网络对解决分类问题传统的思路。in

is the output of the convolutional layer p-1, and the global features generated by this layer will eventually be output to the next convolutional layer or fully connected layer. The remaining fully connected layers and output layers in the UPS method are similar to the last two layers of the UPE method, and will not be repeated here. In the UPS method, each possible value of the attribute is finally scored through the softmax function. The higher score is the unknown value to be predicted. This is the traditional idea of neural networks for solving classification problems.

The key points of the present invention are mainly:

1. Using ego-network, a user attribute prediction algorithm based on convolutional neural network-UPE is proposed, which greatly improves the accuracy of missing attribute prediction of marked nodes. Although convolutional neural networks have been widely used in natural language processing, computer vision and other fields, it is the first attempt in social networks, especially attribute inference.

2. In view of the inability to obtain the attribute information of users in some social networks, a multi-layer neural network-based user missing attribute prediction algorithm-UPS is proposed, which supplements the exceptions of the UPE algorithm.

3. According to the characteristics of social network, the definition of ego-network is put forward; this definition is different from the traditional ego-network, it not only includes the attributes of the central node and the relationship between friends and the attributes of friends and the relationship between friends relationship, and can also include information such as their behavior and the content of their posts. Processing and abstracting these behavioral information into one-dimensional features of users is different from the previous definition of ego-network.

Because the data of social network involves personal privacy, the data set published on the network almost does not exist. The invention uses the crawler technology based on seed URL to crawl 220,000 users on Zhihu, including 15-dimensional attribute information. These attribute information respectively include username, user ID, gender, age, occupation, major, school, etc. By cleaning the data, the data with the user ID and user name removed is finally put into the UPE and the UPS respectively.

In addition, in order to better verify the good performance of the algorithm proposed in the present invention in inferring unknown attributes of users, commonly used machine learning algorithms such as Bayesian (NB), logistic regression (LR), majority voting (MV) are selected. , graph-based semi-supervised (GSSL a label propagation algorithm) and other algorithms. The experimental results are shown in Figure 4, where the numerical value represents the correct rate.

As can be seen from Figure 4, when predicting the gender attribute of marked users, the UPE accuracy rate of the method proposed in this paper can reach 79.11%, which is 5 to 27 percentage points higher than the traditional method. Using the neural network method, the correct rate of UPS is lower than that of UPS, because UPE increases the relationship attribute and adds the user's relationship information. When predicting the user's occupational information, it can be seen from Figure 4 that the correct rate of UPE is significantly higher than the others, reaching 76.28%, while the correct rate of MV (majority voting) is the lowest, which is 32.9%. It can be seen from the experimental results that the convolutional neural network can well extract the potential relationship between the marked node and its friends, so as to better predict the unknown attribute value of the marked node.

The present invention is not limited to the above embodiments, the parameter settings in the neural network, the size of each layer of convolution kernel, and the selection of the activation function (ReLU) are all replaceable, and are selected or trained according to data.

The above embodiments are only used to illustrate the technical solutions of the present invention rather than limit them. Those of ordinary skill in the art can modify or equivalently replace the technical solutions of the present invention without departing from the spirit and scope of the present invention. The scope of protection shall be subject to what is stated in the claims.

Claims (5)

1. a user attribute inference method based on convolutional neural network, is characterized in that, comprises the following steps: 1) According to the attributes and friend relationships of user nodes, establish a self-centered network; 2) using a convolutional neural network to extract the attribute information of the user node in the self-center network and the hidden information contained in the friendship relationship, and use the hidden information to infer the missing attribute of the user; The self-center network is represented by a quintuple G'={V', E', A', B', L}, where V' includes the node information in the self-center network, and the set E' includes all the nodes in the self-center network. The link relationship between nodes, sets A' and B' respectively represent the attribute information and behavior information of the node, the matrix L∈V'×N contains the attribute information and behavior information of the self-center node and its friends, and N is the attribute and behavior data. the sum of dimensions; Step 1) firstly filter the attribute information of the user on the network, and then establish the self-centered network; the filtering includes: a) Filter out all non-Chinese phrases with attributes other than age; b) Filter out nodes whose attribute information is missing beyond the set threshold; The convolutional neural network includes an input layer, a projection layer, a convolution layer, a pooling layer, a fully connected layer and an output layer, and the projection layer converts the user's attribute information and behavior information into vectors; The projection layer converts the user's attribute information and behavior information into vectors, and adopts the following rules for words belonging to the same occupation or profession: i. Create a hash table, take the first word of the profession or occupation as the key, and use the set of all attributes with the first word as the value as the value; then calculate the similarity of other attribute values with the words in the professional and occupation dictionary according to the jara-winkler distance properties, and add attributes with high similarity to the set value; ii. For those attribute values that do not appear in the set value, use word2vec to train word vectors. By calculating the distance between the word vectors, the vectors with high similarity are aggregated together by the KNN algorithm, and the words in the dictionary are passed ID numbers. Correlate to get a digitized vector; The convolutional layer uses the trained weight matrix and bias to map with the feature matrix obtained after the projection layer, and uses Relu as the activation function; the pooling layer uses maximum pooling to retain the most important local features. The output layer uses the softmax classifier to score the possible values of the attribute by using the weight matrix and the bias value, and the attribute value with the highest score is the missing attribute value of the marked user. 2 . The method according to claim 1 , wherein, for social networks in which friend relationships cannot be directly obtained or are difficult to obtain, a neural network is used to classify and predict the missing attributes using only the user's attribute information. 3 . 3. The method according to claim 2, wherein the neural network comprises an input layer, a projection layer, a hidden layer and an output layer; the projection layer converts the user's attribute information and behavior information into vectors; the The hidden layer is two fully connected layers, the first hidden layer contains n*n neurons, n is the sum of the dimensions of the user's attributes and behavior data, and the second hidden layer loses some neurons to prevent overfitting; The output layer uses the softmax classifier to score the possible values of the attribute by using the weight matrix and the bias value, and the attribute value with the highest score is the missing attribute value of the marked user. 4. A device for inferring user attributes based on a convolutional neural network using the method of claim 1, characterized in that, comprising: The self-centered network building module is responsible for establishing a self-centered network according to the attributes and friend relationships of user nodes; The user attribute inference module is responsible for extracting the attribute information of the user node in the self-center network and the hidden information contained in the friend relationship by using a convolutional neural network, and using the hidden information to infer the missing attribute of the user. 5. The device according to claim 4, characterized in that, for social networks that cannot be directly obtained or difficult to obtain for friend relationships, the user attribute inference module adopts a neural network to only use the user's attribute information to carry out the missing attributes. Classification prediction.

Images