CN119378906A

CN119378906A - A task matching method based on user information and task characteristics

Info

Publication number: CN119378906A
Application number: CN202411514124.1A
Authority: CN
Inventors: 刘广澜; 杨雪
Original assignee: Yangyang Beijing Technology Co ltd
Current assignee: Yangyang Beijing Technology Co ltd
Priority date: 2024-10-29
Filing date: 2024-10-29
Publication date: 2025-01-28

Abstract

The present invention relates to the field of task matching technology, and provides a task matching method based on user information and task features, including step 1, collecting and analyzing user information, wherein the user information is not limited to employment information, areas of expertise, experience and price; step 2, parsing task information, wherein the task information is not limited to basic information and additional information; the basic information collection includes but is not limited to task title, task description, task category, budget range, time requirement, task importance, and geographical location requirement; step 3, establishing a multi-dimensional matching model to map user information and task features; step 4, optimizing and sorting matching results using a machine learning algorithm; and step 5, dynamically adjusting matching weights to continuously improve matching accuracy based on historical data. The present invention provides a task matching method based on user information and task features, improves the efficiency and stability of task allocation, and provides personalized task recommendations.

Description

Task matching method based on user information and task characteristics

Technical Field

The invention relates to the technical field of task matching, in particular to a task matching method based on user information and task characteristics.

Background

The task matching platform is an online platform for providing task release and task bearing services, and aims to help a task publisher find a proper task executor and ensure that the task can be completed efficiently. The platform generally has the functions of task release, task classification, user matching, online communication, commission settlement and the like, and provides a convenient task management mode for users.

The conventional task matching platform generally performs simple matching through keywords in task descriptions and user skill labels, performs preliminary screening according to task categories and user proficiency fields, and performs matching according to task budgets and user quotations. However, the current task matching platform has the following technical problems:

1. the matching efficiency is low, and the existing system often adopts simple keyword matching or single-dimension sorting, so that the most suitable task executors cannot be found out from a large number of users. Manual screening is time-consuming and labor-consuming, and is difficult to cope with large-scale task matching requirements.

2. The matching precision is insufficient, and the traditional method is difficult to comprehensively consider the multi-dimensional characteristics (such as skills, experience, price expectation and the like) of the user and the multi-dimensional requirements of the task. Matching in a single dimension (e.g., considering only price or skill labels) often results in unsuitable matching results.

3. The user information is underutilized, and existing systems often ignore implicit features in the user information, such as industry experience reflected by work history. It is difficult to efficiently integrate and utilize dynamic information of users, such as the performance of recently completed tasks.

4. There is a lack of personalized recommendations-no customized task recommendation can be provided according to the user's unique needs and preferences. It is difficult to balance the needs and desires of both the task publisher and the recipient.

Disclosure of Invention

The task matching method based on the user information and the task features mainly solves the technical problems that a task matching platform in the prior art is low in matching efficiency, insufficient in matching precision, insufficient in user information utilization and lack of personalized recommendation, and improves task distribution efficiency and stability and provides personalized task recommendation.

The invention provides a task matching method based on user information and task characteristics, which comprises the following steps:

Step 1, collecting and analyzing user information including, but not limited to job information, direction of expertise, experience, and price;

Step 2, analyzing task information, wherein the task information comprises basic information and additional information, the basic information comprises task titles, task descriptions, task categories, budget ranges, time requirements, task importance and geographic position requirements, the additional information comprises a required skill list, experience level requirements, expected workload, project scale and complexity, accessories and reference materials;

step 3, a multi-dimensional matching model is established, and mapping is carried out on the user information and the task characteristics;

step 4, optimizing and sequencing the matching result by using a machine learning algorithm;

and 5, dynamically adjusting the matching weight, and continuously improving the matching precision according to the historical data.

Further, the manner of collecting and analyzing the user information includes, but is not limited to, user registration, personal data improvement, platform use process, questionnaire, and third party data integration;

the user registers to obtain basic information, identity verification information and education background of the user;

the personal data is perfect, and skill labels, work experiences, professional certificates and work sets of users are obtained;

The platform use process includes, but is not limited to, task history, evaluation feedback, behavioral data, social interactions;

The questionnaire surveys, user surveys are regularly conducted, preference, demand and satisfaction are known, and feedback and suggestion of users on platform functions are collected;

The third party data integration includes, but is not limited to, social media account association, ratings by other specialized platforms, or integration of credit records.

Further, the step 2 includes the following steps 201 to 204:

step 201, performing text preprocessing on task information;

Step 202, performing natural language processing on task information;

Step 203, extracting key information by using a rule base method and/or a machine learning algorithm, and constructing task feature vectors according to the key information;

Step 204, analyzing the task information, and performing one or more of task topic identification, semantic analysis, complexity assessment, similar task matching, priority and urgency analysis, quality control requirements, collaborative demand analysis, risk assessment, normalization and structuring, dynamic updating, and feedback loops.

Further, the step 204 includes the following steps 2041 to 2051:

Step 2041, analyzing task description in the task information, and identifying task subjects;

Step 2042, performing semantic analysis on the task description;

step 2043, performing complexity evaluation on the task information;

step 2044, comparing the task information with a historical task database to identify similar historical tasks;

Step 2045, evaluating the priority according to the time requirement, the budget range and the task importance, identifying keywords and phrases which are in the task description and indicate the emergency degree, and determining the task emergency degree;

step 2046, analyzing terms and standards related to quality in the task description, identifying specific auditing, testing or acceptance standards, and determining task quality control requirements;

step 2047, evaluating whether the task requires team cooperation, and identifying possible subtasks or modules in the task;

Step 2048, identifying potential risk factors in the task description, and evaluating feasibility and potential challenges of the task;

Step 2049, converting the information extracted in steps 2041 to 2042 into a standardized format;

step 2050, acquiring whether the user issues supplementary information and questions in real time, and dynamically updating analysis task information;

step 2051, collecting feedback after the task is completed, and evaluating the accuracy of analysis.

Further, the step 3 includes the following steps 301 to 314:

Step 301, constructing a feature vector of a matching model, wherein the feature vector comprises a user feature vector and a task feature vector, and the user feature vector U and the task feature vector T are respectively defined as follows:

U=[u₁,u₂,...,u_n],T＝[t₁,t₂,...,t_n]

Where n is the number of feature vector dimensions;

Step 302, calculating the similarity between the user feature vector and the task feature vector by using the cosine similarity:

similarity(U,T)=(U·T)/(||U||*||T||);

step 303, assigning a weight to each feature vector, and calculating a weighted matching score:

score(U,T)=Σ(w_i*similarity(u_i,t_i))

Where w _i is the weight of the i-th dimension, Σw _i =1;

Step 304, performing nonlinear conversion on the feature vector:

U'＝f(U)＝[f(u₁),f(u₂),...,f(u_n)]T'＝f(T)＝[f(t₁),f(t₂),...,f(t_n)]

wherein, f can adopt nonlinear activation functions such as ReLU, sigmoid and the like;

Step 305, further processing the converted feature vectors using a multi-layer perceptron (MLP):

H=MLP([U',T'])

wherein the MLP comprises a plurality of fully connected layers and a nonlinear activation function;

at step 306, attention mechanisms are introduced to dynamically adjust the importance of the different features:

A=Attention(H)H'=A⊙H

Wherein, as follows, element-wise multiplication;

step 307, add residual connection to alleviate the difficulty of deep network training:

R=H'+[U',T']

step 308, using an output layer to map the processed features to final matching scores:

score_final=σ(W_out·R+b_out)

wherein σ is a sigmoid activation function;

Step 309, training a matching model using the historical matching data, minimizing a loss function:

L=Σ(y_true-score_final)²+λ||θ||²

where y_true is the actual match result (0 or 1), lambda theta ² is an L2 regularization term;

Step 310, iteratively updating parameters of the matching model using an optimization algorithm;

step 311, integrating a plurality of models with different structures, and adjusting the matching model:

score_ensemble=Σ(w_i*score_model_i)

Where w _i is the weight of each model;

step 312, dynamically adjusting the matching model according to the real-time feedback and the platform state;

step 313, analyze model decision basis, SHAP_values=SHAP (model, [ U, T ])

Step 314, taking into account a plurality of objective functions, finding a balance between the objectives using a multi-objective optimization algorithm (e.g., NSGA-II):

score_multi=[score_accuracy,score_satisfaction,score_efficiency]

Through the above processing of steps 301 to 314, a complex and powerful multidimensional matching model is established, and user information and task characteristics can be mapped.

Further, in step 4, the machine learning algorithm optimizes the matching result using a gradient boosting decision tree algorithm.

Further, the step 5 includes the following steps 501 to 515:

step 501, setting initial weights based on expert knowledge and historical data, and automatically generating initial weight distribution by using a Bayesian optimization method;

Step 502, collecting feedback data of a user;

Step 503, dynamically adjusting by using the time attenuation model;

Step 504, online learning algorithm, update feature weights in real time:

wherein eta is the learning rate and L_t is the loss function at time t;

step 505, dynamically adjusting by using a multi-arm slot machine strategy;

step 506, performing context awareness adjustment;

Step 507, performing differential weight adjustment;

step 508, performing anomaly detection and processing;

step 509, performing an A/B test;

Step 510, performing long-term and short-term memory optimization;

Step 511, performing transfer learning;

step 512, performing federal learning;

Step 513, reinforcement learning optimization;

step 514, performing causal inference;

step 515, automated hyper-parameter tuning is performed.

The task matching method based on the user information and the task characteristics, which is provided by the invention, is applied to a task matching platform, can improve the efficiency and the accuracy of task allocation, and has the following advantages compared with the prior art:

1. High efficiency

1.1, Quick matching, namely, through multidimensional feature vectors and efficient similarity calculation, the system can quickly complete matching in large-scale user and task data.

And 1.2, parallel processing, namely supporting parallel computing by algorithm design, fully utilizing modern computing hardware and further improving the processing speed.

1.3, Incremental update, namely supporting incremental data update without recalculating all data each time, thereby improving the response speed of the system.

1.4, Real-time response capability, namely using INCREMENTAL DECISION TREES to update the model structure once every N samples, adopting an incremental learning algorithm to update the model in real time, and being capable of quickly responding to the changes of user behaviors and preferences.

2. Accuracy of

2.1, Taking into consideration multiple dimensions such as skills, experience, price, time and the like into consideration comprehensively, and providing a more comprehensive matching result. By adopting the Multi-Objective reinforcement learning method, a plurality of Objective functions are optimized by using a Multi-Objective Deep Q-Networks (MODQN), so that a plurality of possible conflict objectives can be optimized simultaneously.

2.2, Semantic understanding, namely more accurate understanding of task description and user skills can be realized through advanced natural language processing technology.

And 2.3, context awareness, namely, taking the time, place and other context information of the task into consideration, and providing matching which is more in line with the actual situation.

And 2.4, robustness and stability, namely integrating a plurality of base models by using a Stacking integration method, reducing the instability of a single model, and having stronger resistance to noise data and abnormal input.

2.5, Interpretive, using decision tree and neural network mixed model, decision=rule_Engine (nn_output), integrating Rule-based system and machine learning model, the matching result has higher interpretive, easy to understand and trust for user.

2.6, Solving the cold start problem, namely quickly adapting to a new scene by using a meta Learning technology, adapting to a new user/task by using a MAML (Model-Agnostic Meta-Learning) algorithm, and effectively processing the matching problem of the new user and the new task.

3. Adaptive property

And 3.1, dynamic weight adjustment, namely automatically adjusting the weight of each dimension by the system according to the actual matching effect through a reinforcement learning algorithm.

3.2, Market adaptation, namely being capable of rapidly adapting to emerging task types and constantly changing user skill structures.

And 3.3, personalized recommendation, namely providing personalized task recommendation according to the historical behaviors and preferences of the user. Providing highly personalized task matching for each user, introducing personalized attention mechanisms, and dynamically adjusting the characteristics for each user.

And 3.4, privacy protection, namely adopting a differential privacy technology to increase noise to sensitive data, and effectively protecting user privacy while improving matching precision.

And 3.5, expandability, namely adopting a distributed computing framework, performing large-scale data processing and model training by using APACHE SPARK, wherein the system framework has good expandability and can process large-scale data and concurrent requests.

And 3.6, continuously learning ability, namely, using a progressive network (Progressive Networks), keeping the prior knowledge and adapting to new tasks, continuously learning from new data, continuously improving the performance, and realizing the life learning of the model.

Drawings

FIG. 1 is a flow chart of an implementation of a task matching method based on user information and task features provided by the invention;

FIG. 2 is a flow chart of the implementation of step 1 provided by the present invention;

FIG. 3 is a flow chart of an implementation of step 2 provided by the present invention;

FIG. 4 is a flow chart of an implementation of step 3 provided by the present invention;

Fig. 5 is a flowchart of the implementation of step 5 provided by the present invention.

Detailed Description

In order to make the technical problems solved by the invention, the technical scheme adopted and the technical effects achieved clearer, the invention is further described in detail below with reference to the accompanying drawings and the embodiments. It is to be understood that the specific embodiments described herein are merely illustrative of the invention and are not limiting thereof. It should be further noted that, for convenience of description, only some, but not all of the matters related to the present invention are shown in the accompanying drawings.

As shown in fig. 1, the task matching method based on the user information and the task characteristics provided by the embodiment of the invention includes the following processes:

Step 1, user information is collected and analyzed including, but not limited to, job information, direction of expertise, experience, and price.

The manner in which the user information is collected and analyzed includes, but is not limited to, user registration, profile improvement, platform usage, questionnaire, and third party data integration.

The user registration can acquire the basic information, the authentication information and the education background of the user. The basic information includes, but is not limited to, a user name, mailbox, phone number, age, gender, etc. The authentication information includes, but is not limited to, real name authentication, identification card number, bank account, etc. The educational background includes, but is not limited to, a top school, a graduation institution, a specialty, and the like.

The personal data is perfect, and the skill label, the work experience, the professional certificate and the work set of the user can be obtained. The skill tags are user-selected or filled in skills that are adept. The work experience includes, but is not limited to, the user's past work experience, job position, time period, etc. The professional certificate is a related qualification certificate, a professional certificate, etc. The work set is a representative work or project case uploaded by the user.

The platform usage includes, but is not limited to, task history, assessment feedback, behavioral data, social interactions. The task history is a record of accepted and completed tasks. The rating feedback is from the rating and rating of the task publisher. The behavior data includes, but is not limited to, login frequency, browsing preferences, search history, and the like. The social interaction is communication records, cooperation conditions and the like with other users.

And the questionnaire surveys are used for regularly carrying out user surveys, knowing preferences, demands and satisfaction and collecting feedback and suggestions of users on platform functions.

The third party data integration includes, but is not limited to, social media account associations (e.g., linkedln), ratings of other specialized platforms, or integration of credit records.

Step 2, analyzing task information, wherein the task information comprises but is not limited to basic information and additional information, the basic information comprises but is not limited to task titles, task descriptions, task categories (such as design, programming, writing and the like), budget ranges, time requirements, task importance and geographic position requirements (if any), and the additional information comprises but is not limited to required skill lists, experience level requirements, expected workload (hours or days), project scale and complexity, accessories and reference materials.

As shown in fig. 2, the step 2 includes the following steps 201 to 204:

In step 201, text preprocessing is performed on the task information.

Text pre-processing can remove special characters and excess blank, correct obvious spelling errors, develop common abbreviations, recognize and standardize technical terms.

Step 202, natural Language Processing (NLP) is performed on the task information.

The task information is split into individual words or phrases.

Labeling the split words or phrases, and distinguishing nouns, verbs, adjectives and the like.

And extracting proper nouns in the specific field from the split words or phrases. Proper nouns such as technical names, tools, platforms, etc.

And carrying out syntactic analysis on the task information, understanding sentence structure, and identifying the relation between the key information.

Step 203, extracting key information by using a rule base method and/or a machine learning algorithm, and constructing task feature vectors according to the key information.

The key information includes, but is not limited to, core skill requirements, work content emphasis, delivery type, time and progress requirements, quality criteria.

The task feature vectors include explicit and implicit requirements.

Step 204, analyzing the task information, and performing one or more of task topic identification, semantic analysis, complexity assessment, similar task matching, priority and urgency analysis, quality control requirements, collaborative demand analysis, risk assessment, normalization and structuring, dynamic updating, and feedback loops. As shown in fig. 3, the step 204 includes the following steps 2041 to 2051:

Step 2041, analyzing task description in the task information, and performing task topic identification.

In the invention, task topic identification is carried out on task description in task information by utilizing an LDA (latent dirichlet allocation) algorithm and the like. Task topics include main topics and sub-topics.

Step 2042, performing semantic analysis on the task description.

In the invention, word embedding technology (such as Word2Vec, BERT) is used for understanding the semantics of task description, calculating the semantic similarity between task description and skill labels and industry terms, and identifying implicit requirements and context information in task description.

Step 2043, complexity evaluation is performed on the task information.

The task complexity is estimated based on factors such as task description, budget scope, time requirements, and the like, and the actual workload required for task completion is estimated by predicting required skill levels and experience requirements using a machine learning model.

Step 2044, comparing the task information with the historical task database, and identifying similar historical tasks.

Identifying similar historical tasks for reference pricing and skill requirements

Step 2045, evaluating the priority according to the time requirement, the budget range and the task importance, identifying keywords and phrases which are in the task description and implicate the degree of urgency, and determining the degree of urgency of the task.

In step 2046, quality related terms and criteria in the task description are analyzed, specific audit, test or acceptance criteria are identified, and task quality control requirements are determined.

In step 2047, the task is evaluated for the need for team collaboration, identifying possible subtasks or modules in the task.

At step 2048, potential risk factors in the task description are identified, and the feasibility and potential challenges of the task are assessed.

Step 2049, converting the information extracted in steps 2041 to 2042 into a standardized format.

In the step, the structured task metadata is generated, so that the subsequent matching and searching are facilitated

Step 2050, obtaining whether the user issues the supplementary information and the query in real time, and dynamically updating the analysis task information.

And updating the task analysis result according to the supplementary information and the answers of the task publishers, and continuously optimizing the analysis algorithm by using a machine learning model.

The present step uses feedback information to continuously improve the analytical algorithm.

And 3, establishing a multi-dimensional matching model, and mapping the user information and the task characteristics. As shown in fig. 4, the step 3 includes the following steps 301 to 314:

in step 301, feature vectors of the matching model are constructed. The feature vectors include user feature vectors and task feature vectors.

The user feature vector U and the task feature vector T are defined as follows:

U=[u₁,u₂,...,u_n],T＝[t₁,t₂,...,t_n]

Where n is the number of feature vector dimensions, task feature vectors include, but are not limited to, skill matching, price expectations, time availability, experience level, and the like.

similarity(U,T)=(U·T)/(||U||*||T||)

score(U,T)=Σ(w_i*similarity(u_i,t_i))

Where w _i is the weight of the i-th dimension, Σw _i =1.

Step 304, performing nonlinear conversion on the feature vector:

Wherein, f can adopt nonlinear activation functions such as ReLU, sigmoid and the like. This step can capture complex relationships between features. The enhancement model captures the ability of complex patterns.

H=MLP([U',T'])

Wherein the MLP comprises a plurality of fully connected layers and nonlinear activation functions for learning deep relationships between user features and task features.

A=Attention(H)H'=A⊙H

Wherein, as indicated by element-wise multiplication. The attention mechanism can adaptively adjust the feature weights according to specific user-task pairs.

R=H'+[U',T']

this step helps to preserve the original characteristic information and prevents information from being lost in the multi-layer delivery.

score_final=σ(W_out·R+b_out)

Where σ is a sigmoid activation function that compresses the output to between 0-1, indicating the degree of matching.

L=Σ(y_true-score_final)²+λ||θ||²

Where y_true is the actual matching result (0 or 1), λ θ ² is the L2 regularization term, preventing overfitting.

In step 310, the parameters of the matching model are iteratively updated using an optimization algorithm.

The optimization algorithm is, for example, a gradient descent method. And updating parameters of the matching model, and continuously improving the matching accuracy.

Step 311, integrating a plurality of models with different structures (such as CNN and RNN-based variants), adjusting the matching model, and further improving the robustness:

score_ensemble=Σ(w_i*score_model_i)

where w _i is the weight of each model, which can be determined by validation set performance.

The above steps mainly used the following models a and B. The model A is a CNN-based matching model, namely, a convolutional neural network is used for processing the characteristics, so that the local relation among the characteristics can be captured, and the model A is suitable for processing the characteristic modes related to the positions. The model B is a matching model based on Attention, is a pure Attention mechanism architecture, and is used for better processing long-distance dependence among features and dynamically adjusting feature importance.

Step 312, dynamically adjusting the matching model according to the real-time feedback and the platform state.

Model parameters are updated using a line learning algorithm. The matching model is periodically retrained to accommodate changes in user behavior and task characteristics.

Step 313, analyze model decision basis, SHAP_values=SHAP (model, [ U, T ])

The step can analyze model decision basis by using SHAP value and other technologies to provide interpretable matching reasons. This helps to increase the user's confidence in the matching results and provides insight into further optimization.

score_multi=[score_accuracy,score_satisfaction,score_efficiency]

such as matching accuracy, user satisfaction, platform efficiency, etc.

Through the processing from step 301 to step 314, the invention establishes a complex and powerful multidimensional matching model, and can effectively map user information and task characteristics. The model not only considers the linear and nonlinear relations between features, but also introduces advanced technologies such as deep learning, attention mechanisms and the like to capture finer matching patterns. Meanwhile, through model integration, dynamic adjustment and interpretability analysis, the robustness, adaptability and transparency of the model are ensured. Finally, this mapping process can generate a comprehensive matching score for each user-task pair, providing a reliable basis for subsequent task assignments and recommendations.

And 4, optimizing and sequencing the matching result by using a machine learning algorithm.

The machine learning algorithm of the present invention uses a Gradient Boosting Decision Tree (GBDT) algorithm to optimize the matching results. The GBDT model can be expressed as:

F(U,T)=Σf_k(U,T)

where f _k is the kth decision tree.

The model is trained by minimizing the following loss functions:

L=Σl(y_i,F(U_i,T_i))+ΣΩ(f_k)

where l is the loss function (e.g., log loss), y _i is the true match result, and Ω (f _k) is the regularization term.

The gradient lifting decision tree (GBDT) algorithm is a powerful machine learning technology, and can remarkably improve matching accuracy. GBDT is an ensemble learning method that gradually optimizes the predicted outcome by building multiple decision trees. Each new tree is generated to correct the residual of the previous tree. The final model is a weighted sum of all trees, F (U, T) =Σf _k (U, T), where U is the user feature vector, T is the task feature vector, and F _k is the kth decision tree.

Model training process of gradient lifting decision tree algorithm:

a) Initialization F ₀(U,T)＝arg min_γΣl(y_i, gamma

The model is initialized with a constant value, typically the average of the target variables.

B) For each round of iterations m=1, 2, M:

Calculate negative gradient (pseudo residual):

fitting a regression tree h _m (U, T) to the residual r _im;

calculating a leaf node region R _jm,j＝1,2,...,J_m;

calculating an optimal value of gamma _jm＝arg min_γΣl(y_i,F_m-1(U_i,T_i) +gamma for each leaf node;

Updating the model F _m(U,T)＝F_m-1(U,T)+ν·Σγ_jmI((U,T)∈R_jm) where v is the learning rate, for controlling the overfit.

In the matching problem, to prevent model overfitting, the present invention employs a regularization term Ω (f _k)＝γT+λ||w||², where T represents the number of leaf nodes of the tree, w is the weight vector of the leaf nodes, and γ and λ are hyper-parameters used to control the complexity of the model.

This may be referred to as a "dual regularization loss function of the Tree structure" (Tree-based Dual Regularization Loss Function) because it constrains both the structure of the Tree (node number T) and the leaf weights (w).

Feature importance GBDT can naturally provide a feature importance measure of the cumulative contribution of features to model prediction improvement by the frequency at which features are selected as split points in the tree.

Processing class features by single-hot encoding class features or using special splitting strategies (e.g. Categorical Embedding)

Hyper-parameter optimization, which is to adjust key hyper-parameters by using a grid search method or a Bayesian optimization method, wherein the key hyper-parameters comprise, but are not limited to, the number of trees, the maximum depth of the trees, the learning rate, the minimum leaf node sample number and L1 and L2 regularization parameters.

Model integration multiple GBDT models may be trained and integrated. Bagging, training a model on different subsets of data. Boosting, using different initialization or super parameter settings.

Online learning and incremental updating-the model is updated step by step using forgetting factors, and the partial tree is periodically retrained to accommodate new data distributions.

Match score calculation the final match score can be expressed as:

score(U,T)=σ(F(U,T))

Where σ is a sigmoid function, compressing the output to between 0-1.

Sequencing the results, namely sequencing the user-task pairs according to the calculated matching score:

ranked_matches＝sort([(U_i,T_i,score(U_i,T_i))for all i],key＝lambda x:x[2],reverse＝True)

post-processing, applying business rules (e.g., considering user load, task urgency, etc.), introducing diversity (e.g., using maximum marginal correlation algorithms).

Interpretive-the decision basis for each match is interpreted using the SHAP (SHAPLEY ADDITIVE exPlanations) values.

Performance monitoring, namely tracking offline evaluation indexes (such as AUC and NDCG) and monitoring online indexes (such as click rate, completion rate and user satisfaction).

A/B testing-comparing GBDT actual effects with other algorithms (e.g., neural networks) by online experiments.

Through the detailed GBDT optimization process, the invention can fully utilize the multidimensional characteristics of users and tasks, learn complex nonlinear relations and generate high-quality matching results. The method can not only improve the accuracy of matching, but also provide interpretable results, and is helpful for continuously improving the performance of a matching system.

And 5, dynamically adjusting the matching weight, and continuously improving the matching precision according to the historical data. As shown in fig. 5, the step 5 includes the following steps 501 to 515:

step 501, setting initial weights based on expert knowledge and historical data, and automatically generating initial weight distribution by using a Bayesian optimization method.

Step 502, collecting feedback data of a user.

Feedback data includes, but is not limited to, recording user behavior to accept/reject tasks, collecting task completion and quality scores, and tracking user satisfaction survey results.

In step 503, dynamic adjustment is performed using the time-decay model.

Introducing a time attenuation factor w (t) =w ₀. Exp (- λt)

Older matching results are given lower weight, focusing more on recent performance.

Step 504, online learning algorithm, update feature weights in real time:

Where η is the learning rate and l_t is the loss function at time t.

The online learning algorithm is, for example, the Follow-the-Regularized-Leader (FTRL) algorithm.

Step 505, dynamically adjusting by using a multi-arm slot machine strategy.

Specifically, each matching strategy is considered as an arm. The Thompson Sampling algorithm was used to balance exploration and utilization. The frequency of use of the different matching strategies is dynamically adjusted.

In step 506, context awareness adjustment is performed.

Specifically, the time, place, user state, and other contextual information are considered. Context dependencies are captured using a Conditional Random Field (CRF) model.

In step 507, a differential weight adjustment is performed.

Specifically, independent weight adjustment strategies are set for different user groups and task types. Multiple related weight adjustment tasks are simultaneously optimized using a multi-task learning framework.

Step 508, performing anomaly detection and processing.

Specifically, an isolated forest algorithm is used to detect abnormal matching results. And carrying out special processing on the abnormal data to prevent the error data from affecting the whole model.

Step 509, an A/B test is performed.

Specifically, small-scale a/B testing was continued, and different weight adjustment strategies were evaluated. The multi-arm slot machine algorithm is used to automatically distribute traffic to the best performing strategy.

Step 510, long-term memory optimization is performed.

Specifically, the combination of LSTM networks captures both long-term preferences and short-term interest changes of users. The feature weights are dynamically adjusted to balance long term stability with short term flexibility.

In step 511, the migration learning is performed.

Specifically, new task types are quickly adapted using weight knowledge learned in similar fields. The domain adaptation technique is used to handle the distribution differences between the different task domains.

Step 512, federal learning.

Specifically, on the premise of protecting the privacy of the user, the data on the distributed equipment is utilized to update the model. Model updates from multiple users are consolidated using a secure syndication protocol.

Step 513, reinforcement learning optimization.

Specifically, the matching process is modeled as a Markov Decision Process (MDP), and the optimal weight adjustment strategy is learned using a Deep Q Network (DQN).

Step 514, make causal inference.

Specifically, the causal effect of weight adjustment is evaluated using a counter-facts analysis. The bias is reduced using a Double machine learning (Double MACHINE LEARNING) method.

Step 515, automated hyper-parameter tuning is performed.

Specifically, the super parameters such as the learning rate, regularization parameters and the like are automatically adjusted by using Bayesian optimization. Periodically re-evaluating and updating the super parameters to adapt to the change of the data distribution.

It should be noted that the above embodiments are merely for illustrating the technical solution of the present invention, and not for limiting the same, and although the present invention has been described in detail with reference to the above embodiments, it should be understood by those skilled in the art that the technical solution described in the above embodiments may be modified or some or all of the technical features may be equivalently replaced, without departing from the spirit of the technical solution of the present invention.

Claims

1. The task matching method based on the user information and the task characteristics is characterized by comprising the following steps of:

2. The task matching method based on user information and task features according to claim 1, wherein the manner of collecting and analyzing user information includes, but is not limited to, user registration, personal data improvement, platform use process, questionnaire, third party data integration;

3. The task matching method based on user information and task characteristics according to claim 1, wherein the step 2 includes the following steps 201 to 204:

step 201, performing text preprocessing on task information;

Step 202, performing natural language processing on task information;

4. The task matching method based on user information and task characteristics according to claim 3, wherein the step 204 includes the following steps 2041 to 2051:

Step 2042, performing semantic analysis on the task description;

step 2043, performing complexity evaluation on the task information;

5. The task matching method based on user information and task characteristics according to claim 4, wherein the step 3 includes the following steps 301 to 314:

U=[u₁,u₂,...,u_n],T＝[t₁,t₂,...,t_n]

Where n is the number of feature vector dimensions;

similarity(U,T)=(U·T)/(||U||*||T||);

score(U,T)=Σ(w_i*similarity(u_i,t_i))

Where w _i is the weight of the i-th dimension, Σw _i =1;

Step 304, performing nonlinear conversion on the feature vector:

H=MLP([U',T'])

A=Attention(H)H'=A⊙H

Wherein, as follows, element-wise multiplication;

R=H'+[U',T']

score_final=σ(W_out·R+b_out)

wherein σ is a sigmoid activation function;

L=Σ(y_true-score_final)²+λ||θ||²

score_ensemble=Σ(w_i*score_model_i)

Where w _i is the weight of each model;

step 313, analyze model decision basis, SHAP_values=SHAP (model, [ U, T ])

score_multi=[score_accuracy,score_satisfaction,score_efficiency]

6. The task matching method based on user information and task features according to claim 5, wherein in step 4, the machine learning algorithm optimizes the matching result using a gradient boosting decision tree algorithm.

7. The task matching method based on user information and task characteristics according to claim 6, wherein the step 5 includes the following steps 501 to 515:

Step 502, collecting feedback data of a user;

Step 503, dynamically adjusting by using the time attenuation model;

Step 504, online learning algorithm, update feature weights in real time:

wherein eta is the learning rate and L_t is the loss function at time t;

step 505, dynamically adjusting by using a multi-arm slot machine strategy;

step 506, performing context awareness adjustment;

Step 507, performing differential weight adjustment;

step 508, performing anomaly detection and processing;

step 509, performing an A/B test;

Step 510, performing long-term and short-term memory optimization;

Step 511, performing transfer learning;

step 512, performing federal learning;

Step 513, reinforcement learning optimization;

step 514, performing causal inference;

step 515, automated hyper-parameter tuning is performed.