A recommendation method and system for learning a knowledge graph by fusing interactive attention
By employing a multi-task learning shared module and feature attention cross-compression technology, the data sparsity and cold start problems in traditional recommendation methods are solved, thereby improving the accuracy and effectiveness of recommendations.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HUBEI UNIV OF TECH
- Filing Date
- 2023-05-25
- Publication Date
- 2026-05-29
AI Technical Summary
Traditional recommendation methods perform poorly when faced with data sparsity and cold start problems, and struggle to effectively utilize contextual information and address insufficient interaction.
A multi-task learning sharing module is introduced, which trains the knowledge graph and recommendation model through interaction. It combines user-item interaction information and knowledge graph data, and uses a feature attention cross-compression module and a multi-layer neural network for feature extraction and prediction.
It effectively alleviates the problems of data sparsity and cold start, improves the accuracy and effectiveness of recommendations, and can quickly filter out useful information from a large amount of data.
Smart Images
Figure CN116701611B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence technology, and more specifically, relates to a recommendation method and system for a learning knowledge graph that integrates interactive attention. Background Technology
[0002] Before the advent of the era of big data and artificial intelligence, traditional recommendation models were generally content-based, collaborative filtering, and model-based recommendations. However, with the explosive growth of data volume and the increasing diversity of data types, the data sparsity and cold start problems exposed by traditional recommendation methods have become increasingly apparent, rendering them inadequate. With the rapid development of deep learning, deep learning algorithms have been successfully integrated into recommendation models. Recommendation models incorporating deep learning have stronger learning and expressive capabilities, attracting attention both domestically and internationally. Summary of the Invention
[0003] This invention addresses the technical problems existing in the prior art by providing a recommendation method and system that integrates interactive attention-based learning knowledge graphs. Compared to traditional knowledge graph-enhanced recommendation methods, this invention introduces a multi-task learning and sharing module to interactively train the knowledge graph and recommendations, thus alleviating the data sparsity and cold-start problems in traditional recommendation methods.
[0004] According to a first aspect of the present invention, a recommendation method for a learning knowledge graph that integrates interactive attention is provided, including a multi-task learning sharing module, comprising the following steps:
[0005] Step 1: Obtain user-item interaction information from the recommendation dataset and knowledge graph information from the knowledge graph dataset; construct a user-item interaction matrix based on the historical records of user ratings of items, and construct a knowledge graph after establishing a mapping between the entities in the knowledge graph dataset and the recommendation dataset; wherein, the user-item interaction information includes the set of users and items, the historical records of user ratings of items, and the knowledge graph is composed of head entity-relationship-tail entity triples;
[0006] Step 2: Input the item features and head entity features obtained from the user-item interaction information into the multi-task learning and sharing module;
[0007] Step 3: Based on the multi-task learning sharing module, the knowledge graph module is trained through alternating training, and the scoring function of the knowledge graph triples is obtained through the similarity function;
[0008] Step 4: Based on the multi-task learning sharing module, the recommendation module is trained through alternating training, and finally the prediction probability of user participation in the project is obtained through the prediction function.
[0009] Based on the above technical solution, the present invention can also be improved as follows.
[0010] Optionally, the multi-task learning sharing module consists of multiple feature attention cross-compression modules. The feature attention cross-compression module includes a feature interaction attention unit and a cross-compression unit. The item features and head entity features are learned by the feature interaction attention unit to obtain item features and entity features that fuse contextual information, which prepares for interactive training of the recommendation module and the knowledge graph module.
[0011] Optionally, the recommendation module includes multiple feature extraction layers and a prediction layer connected in sequence.
[0012] Optionally, inputting the item features and head entity features into the multi-task learning and sharing module includes:
[0013] Step 2.1: For each input candidate item and head entity vector, multiply it with the trainable parameter matrix to obtain the query vector and address vector of each vector;
[0014] Step 2.2: Calculate the similarity score between the query vector and the address vector, and then apply a normalized exponential function to the similarity score to obtain their respective weight coefficient scores;
[0015] Step 2.3: Perform a weighted summation of the weight coefficient scores with the item vector or the head entity vector to obtain the item self-attention score and the head entity self-attention score, respectively.
[0016] Step 2.4: Perform a dot product between the item self-attention score and the head entity self-attention score and their self-attention scores processed by the activation function;
[0017] Step 2.5: Add the dot product value to its initial feature vector to obtain the item vector and head entity vector respectively;
[0018] Step 2.6: The item vector and the head entity vector are cross-multiplied in the cross-compression unit to obtain the cross-multiplication matrix;
[0019] Step 2.7: Transpose the cross-multiplication matrix to obtain the transposed cross-multiplication matrix;
[0020] Step 2.8: Compress the cross-formed feature matrix and transpose matrix, and output the feature vectors of the next layer of items and entities.
[0021] Optionally, in step 2.2, a normalized exponential function is applied to the similarity scores to obtain their respective weight coefficient scores, which are expressed as follows:
[0022]
[0023] Where score1 and score2 are the weight coefficient scores of the item and the entity, respectively; d k These are query vectors Q1, Q2, and address vector K1. T K2 T The vector dimension of a matrix.
[0024] Optionally, in step 2.6, the item feature vector and the head entity feature vector are cross-multiplied in the cross-compression unit to obtain a cross-multiplication matrix. By calculating the outer product of the item vector and the head entity vector, a d×d dimensional cross-multiplication matrix is constructed, represented as:
[0025]
[0026] Among them, B L ∈R dxd is the L-layer feature cross-multiplication matrix, and d is the dimension of the embedding vector. It is B L The interaction multiplication of each feature of the item v and its associated head entity h.
[0027] Optionally, in step 2.8, the feature vectors of the next layer of items and entities are calculated and output as follows:
[0028]
[0029]
[0030] in, and It represents the weight vector and bias term of the trainable compressed unit in the Lth layer for the item vector v and the head entity vector h.
[0031] Optionally, in step 3, the scoring function for obtaining the knowledge graph triples through the similarity function includes:
[0032] Step 3.1: For a given triplet, the head entity learns the shared module through a multi-task learning process to obtain the high-order features of the head entity;
[0033] Step 3.2: Extract the latent cross features of relational entities using a multi-layer neural network;
[0034] Step 3.3: Use the knowledge graph embedding function to predict the tail entity's prediction vector by embedding the head entity features and relation features into the knowledge graph.
[0035] Step 3.4: Obtain the score function for triples through the similarity function.
[0036] Optionally, in step 4, the recommendation module, based on the multi-task learning and sharing module, trains itself through alternating training, and finally obtains the final predicted probability of user participation in the project through a prediction function, including:
[0037] Step 4.1: For the user's initial feature vector, after feature embedding, the user's hidden layer compressed features are extracted through an L-layer multi-layer neural network.
[0038] Step 4.2: For the original input vector of the item, the latent features of the item are calculated through the multi-task learning shared module;
[0039] Step 4.3: Calculate the final predicted probability of user participation in the project using the prediction function.
[0040] According to a second aspect of the present invention, a recommendation system integrating a learning knowledge graph with interactive attention is provided, comprising:
[0041] The data acquisition module is used to obtain user-item interaction information from the recommendation dataset and knowledge graph information from the knowledge graph dataset. Based on the historical records of user ratings of items, it constructs a user-item interaction matrix. After establishing a mapping between entities in the knowledge graph dataset and the recommendation dataset, it constructs a knowledge graph. The user-item interaction information includes the set of users and items, the historical records of user ratings of items, and the knowledge graph is composed of head entity-relationship-tail entity triples.
[0042] The interactive training preparation module is used to input the item features and head entity features obtained from user-item interaction information into the multi-task learning sharing module;
[0043] The knowledge graph training module is used to train the knowledge graph module based on the multi-task learning sharing module through alternating training, and obtains the score function of triples through the similarity function;
[0044] The recommendation module training module is used to train the recommendation module through alternating training based on the multi-task learning shared module, and finally obtains the final predicted probability of user participation in the project through the prediction function.
[0045] The technical effects and advantages of this invention are as follows:
[0046] This invention combines knowledge graphs with recommendation methods, solving the problem of inaccurate recommendation results caused by the sparsity of interactive data in traditional recommendation methods. It can quickly help users filter useful information from a huge amount of information, effectively improving the accuracy of recommendations.
[0047] This invention uses a multi-task learning and sharing module to jointly train a knowledge graph and a recommendation module. While solving the data sparsity problem of traditional recommendation methods, it also effectively avoids the cold start problem; it can effectively address the issues of failing to utilize contextual information and insufficient interaction, thereby improving the actual recommendation performance.
[0048] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description
[0049] Figure 1 A model architecture diagram of a recommendation method for learning knowledge graphs that integrates interactive attention, provided in an embodiment of the present invention;
[0050] Figure 2 This is a model diagram of a feature attention cross-compression unit for a recommendation method that integrates interactive attention to a learning knowledge graph, provided in an embodiment of the present invention. Detailed Implementation
[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0052] Understandably, given the deficiencies in the background technology, this embodiment of the invention proposes a recommendation method for learning knowledge graphs that integrates interactive attention. This method includes the following steps:
[0053] Step 1: This step involves obtaining user-item interaction information from the recommendation dataset and knowledge graph data to acquire knowledge graph information. The user-item interaction information includes sets of users and items, and historical records of user ratings of items. Here, U = {u1, u2, ..., um} and V = {v1, v2, ..., vn} represent the sets of users and items, respectively. Based on the historical records of user ratings of items, a user-item interaction matrix Y = {y...} is constructed. uv |u∈U,v∈V}, where y uv A value of 1 indicates an interaction between user u and item v, such as clicking, viewing, or browsing. Where y uvA value of 0 indicates that there is no interaction between user u and item v. Furthermore, after establishing a mapping between entities in the knowledge graph dataset and the recommendation dataset, a knowledge graph G is constructed. Knowledge graph G consists of a large number of entity-relation-entity triples (h, r, t). Here, h∈E, r∈R, and t∈E represent the head entity, relation, and tail entity of the knowledge graph, respectively, where E and R represent the sets of entities and relations in knowledge graph G.
[0054] Step 2: Input the item features and head entity features obtained from the user-item interaction information into the multi-task learning and sharing module.
[0055] like Figure 2 As shown, the multi-task learning sharing module consists of L feature attention cross-compression modules. Each feature attention cross-compression module includes a feature interaction attention unit and a cross-compression unit. The feature interaction attention unit learns both item features and head entity features, distinguishes the importance of different cross features, and learns the contextual semantic information of the head entity, enhancing the model's interpretability. The cross-compression unit captures the connectivity of higher-order cross features between item features and head entity features.
[0056] It should be noted that the L feature attention cross-compression modules can mean one or more, i.e., one or more. Step 2 specifically includes:
[0057] Step 2.1: For each input candidate item and head entity feature vector, represented as v and h, use trainable parameter matrices respectively. and Multiplying it, it is expressed as:
[0058]
[0059]
[0060] Where Q1, K1, and V1 are trainable matrices of the item vectors, representing the query vector, address vector, and value vector matrices corresponding to the item vectors; Q2, K2, and V2 are trainable matrices of the head entity vectors, representing the query vector, address vector, and value vector matrices corresponding to the head entity vectors.
[0061] Step 2.2: Using the query vector and address vector from Step 2.1, calculate the correlation between the query vector and the address vector, i.e., the similarity score between the two vectors. Then, apply a normalized exponential function to the similarity score to obtain their respective weight coefficient scores, which are expressed as follows:
[0062]
[0063] Where, d kThese are query vectors Q1, Q2, and address vector K1. T K2 T The vector dimension of the matrix; score1 and score2 are the weight coefficient scores of the item and the entity, respectively.
[0064] Step 2.3: The weight coefficient scores obtained in Step 2.2 are weighted and summed with the item vector h or the head entity vector v respectively to obtain the final item self-attention score Atten1 and head entity self-attention score Atten2, which are expressed as follows:
[0065] Atten1 = score1·h
[0066] Atten2 = score2·v
[0067] Step 2.4: Perform a dot product between the self-attention score obtained in Step 2.3 and the self-attention score processed by the sigmoid activation function to obtain the item vector v1 and the head entity vector h1, respectively. Here, the normalized exponential function acts as a gating mechanism, which can be used to select more important features from the features. It is expressed as:
[0068] h1 = Atten1·sigmoid(Atten1)
[0069] v1 = Atten2·sigmoid(Atten2)
[0070] Step 2.5: Add the item vector and head entity vector obtained in Step 2.4 to their initial feature vectors respectively, finally obtaining the item vector v and head entity vector h. This can be represented as:
[0071] v = add(Att1, v)
[0072] h = add(Att2, h)
[0073] Step 2.6: The head entity vector and item vector obtained in Step 2.5 are then processed through a cross-compression unit. In the cross-compression layer, the item feature vector v and the head entity feature vector h obtained through the feature interaction attention unit above are subjected to feature cross-compression. Let the item vector in layer L be... Head entity vector Then the cross-multiplication matrix B of the Lth layer L The calculation is expressed as:
[0074]
[0075] Among them B L ∈R dxdis the feature cross-multiplication matrix of the Lth layer, and d is the dimension of the embedding vector. It is B L The interaction multiplication of each feature of the item v and its associated head entity h.
[0076] Step 2.7: Multiply the cross-multiplication matrix B from Step 2.6. L Transpose to obtain the transposed cross-multiplication matrix Represented as:
[0077]
[0078] Step 2.8: Multiply matrix B L Cross-multiplication matrix with transpose By restoring the dimension of the multiplication matrix from d×d to d×1 using the weight vectors, we obtain the item vector v of the (L+1)th layer. L+1 and head entity vector h L+1 It is represented as:
[0079]
[0080]
[0081] in, and It represents the weight vector and bias term of the trainable compressed unit in the Lth layer for the item vector v and the head entity vector h.
[0082] Step 3: Train the knowledge graph module using an alternating training method based on the multi-task learning sharing module.
[0083] In this embodiment, the specific implementation includes the following sub-steps:
[0084] Step 3.1: For the relation entity r, use a multi-layer neural network to extract its latent cross features, which are represented as follows:
[0085] rL=M(M...M(r))=M L (r)
[0086] Where rL represents the features of the relational entity at the Lth level, M L () represents a fully connected layer function.
[0087] Step 3.2: Extract the latent cross features of relational entities using a multi-layer neural network;
[0088] It is represented as:
[0089] rL=M(M...M(r))=M L (r)
[0090] Step 3.3: First, concatenate the head entity features and relation features of the Lth layer, and then use a multi-layer neural network to obtain the predicted vector of the tail entity t. It is represented as:
[0091]
[0092] Step 3.4: Then, obtain the score function for the triples using the similarity function, where the similarity function is t and After performing the inner product, it is then passed through a sigmoid function, which is expressed as:
[0093]
[0094] Here, score(h,r,t) is the similarity score, and σ is the activation function.
[0095] Step 4: Based on the multi-task learning sharing module, the recommendation module is trained through alternating training, and finally the prediction probability of user participation in the project is obtained through the prediction function.
[0096] It should be noted that the recommendation module contains L sequentially connected feature extraction layers and one prediction layer. The L feature extraction layers can be represented as multiple layers.
[0097] In this embodiment, the recommendation module, which is trained by the multi-task learning and sharing module through alternating training, and finally obtains the final predicted probability of user participation in the project through the prediction function, specifically includes the following sub-steps:
[0098] Step 4.1: For the initial feature vector of user u, after feature embedding, the compressed features uL of the user's hidden layer are extracted through an L-layer neural network, which is represented as:
[0099] uL=M(M(...M(u)))=M L (u)
[0100] Step 4.2: For the original input vector v of item v, the latent features of the item in the final output of layer L are calculated through the multi-task learning shared module. It is represented as:
[0101]
[0102] in, and It represents the trainable compression unit weight vector and bias term when the Lth layer performs compression operation on the item vector v.
[0103] Step 4.3: Calculate the final predicted probability of user u participating in item v using the prediction function.
[0104] It is represented as:
[0105]
[0106] It is important to note that the prediction function here is the inner product of the user vector and the item vector, which is then passed through a sigmoid function to obtain the predicted score.
[0107] In summary, compared to traditional knowledge graph-enhanced recommendation methods, this invention introduces a multi-task learning and sharing module to interactively train the knowledge graph and recommendation module, alleviating the data sparsity and cold-start problems in traditional recommendation methods. Firstly, in the multi-task learning module, item features and head entity features are processed through a feature interaction attention unit to obtain interacted features, thus learning finer-grained features. The item features and head entity features processed by the interaction attention module are then alternately trained on the knowledge graph and recommendation module through a cross-compression unit. This invention can fully utilize the information contained in the knowledge graph and recommendation modules, effectively alleviating the problems of failing to utilize contextual information and data sparsity, thereby improving the actual recommendation performance.
[0108] According to a second aspect of the present invention, embodiments of the present invention also provide a recommendation system that integrates a learning knowledge graph with interactive attention, comprising:
[0109] The data acquisition module is used to obtain user-item interaction information from the recommendation dataset and knowledge graph information from the knowledge graph dataset. Based on the historical records of user ratings of items, it constructs a user-item interaction matrix. After establishing a mapping between entities in the knowledge graph dataset and the recommendation dataset, it constructs a knowledge graph. The user-item interaction information includes the set of users and items, the historical records of user ratings of items, and the knowledge graph is composed of head entity-relationship-tail entity triples.
[0110] The interactive training preparation module is used to input the item features obtained from user-item interaction information and the head entity features of the knowledge graph into the multi-task learning and sharing module.
[0111] The knowledge graph training module is used to train the knowledge graph module based on the multi-task learning sharing module through alternating training, and obtains the score function of triples through the similarity function;
[0112] The recommendation module training module is used to train the recommendation module through alternating training based on the multi-task learning shared module, and finally obtains the final predicted probability of user participation in the project through the prediction function.
[0113] It is understood that the recommendation system for learning knowledge graphs that integrates interactive attention provided by this invention corresponds to the recommendation method for learning knowledge graphs that integrates interactive attention provided in the foregoing embodiments. The relevant technical features of the recommendation system for learning knowledge graphs that integrates interactive attention can be referred to the relevant technical features of the recommendation method for learning knowledge graphs that integrates interactive attention, and will not be repeated here.
[0114] It should be understood that any parts not described in detail in this specification belong to the prior art.
[0115] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
[0116] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A recommendation method for a learning knowledge graph that integrates interactive attention, characterized in that, Includes the following steps: Step 1: Obtain user-item interaction information from the recommendation dataset and knowledge graph information from the knowledge graph dataset; Based on the historical records of user-rated items, a user-item interaction matrix is constructed. After establishing a mapping between entities in the knowledge graph dataset and the recommendation dataset, a knowledge graph is constructed. The user-item interaction information includes the set of users and items, and the historical records of user-rated items. The knowledge graph consists of head entity-relationship-tail entity triples. Step 2: Input the item features obtained from user-item interaction information and the head entity features from the knowledge graph into the multi-task learning and sharing module; the multi-task learning and sharing module consists of multiple feature attention cross-compression modules, each including a feature interaction attention unit and a cross-compression unit. The item features and head entity features are learned by the interaction attention unit to obtain item features and entity features fused with contextual information, preparing for interactive training of the recommendation module and the knowledge graph module; including: Step 2.1: For each input candidate item and head entity vector, multiply it with the trainable parameter matrix to obtain the query vector and address vector of each vector; Step 2.2: Calculate the similarity score between the query vector and the address vector, and then apply a normalized exponential function to the similarity score to obtain their respective weight coefficient scores; Step 2.3: Perform a weighted summation of the weight coefficient scores with the item vector or the head entity vector to obtain the item self-attention score and the head entity self-attention score, respectively. Step 2.4: Perform a dot product between the item self-attention score and the head entity self-attention score and their self-attention scores processed by the activation function; Step 2.5: Add the dot product value to its initial feature vector to obtain the item vector and head entity vector respectively; Step 2.6: The item vector and the head entity vector are first cross-multiplied in the cross-compression layer to obtain the cross-multiplication matrix; Step 2.7: Transpose the cross-multiplication matrix to obtain the transposed cross-multiplication matrix; Step 2.8: Compress the cross-cutting features and calculate the feature vectors of the items and head entities in the next layer; Step 3: Based on the multi-task learning sharing module, the knowledge graph module is trained through alternating training, and the scoring function of the knowledge graph triples is obtained through the similarity function; Step 4: Based on the multi-task learning sharing module, the recommendation module is trained through alternating training, and finally the prediction probability of user participation in the project is obtained through the prediction function.
2. The recommendation method for a learning knowledge graph that integrates interactive attention according to claim 1, characterized in that, The recommendation module comprises multiple feature extraction layers connected in sequence and a prediction layer.
3. The recommendation method for a learning knowledge graph that integrates interactive attention according to claim 1, characterized in that, In step 2.2, a normalized exponential function is applied to the similarity scores to obtain their respective weight coefficient scores, expressed as follows: , in, and These are the weighting coefficient scores for items and the weighting coefficient scores for entities, respectively. It is a query vector , and address vector , The vector dimension of a matrix.
4. The recommendation method for a learning knowledge graph that integrates interactive attention according to claim 1, characterized in that, In step 2.6, the item vector and head entity vector undergo feature cross-compression in the cross-compression unit to obtain a cross-multiplication matrix. By calculating the outer product of the item vector and head entity vector, a cross-multiplication matrix is constructed. A cross-multiplication matrix of dimension 1 is represented as: in, is the feature cross-multiplication matrix of the Lth layer, and d is the dimension of the embedding vector. yes The interaction multiplication of each feature of the item v and its associated head entity h.
5. The recommendation method for a learning knowledge graph that integrates interactive attention according to claim 1, characterized in that, In step 2.8, the feature vectors of the items and head entities in the next layer are calculated and output as follows: in, , , and , Let represent the weight vector and bias term of the trainable compressed unit in the Lth layer for the item vector v and the head entity vector h, respectively.
6. The recommendation method for a learning knowledge graph that integrates interactive attention according to claim 1, characterized in that, In step 3, the scoring function for obtaining knowledge graph triples through the similarity function includes: Step 3.1: For a given triplet, the head entity learns the shared module through a multi-task learning process to obtain the high-order features of the head entity; Step 3.2: Extract the latent cross features of relational entities using a multi-layer neural network; Step 3.3: Use the knowledge graph embedding function to predict the tail entity's prediction vector by embedding the head entity features and relation features into the knowledge graph. Step 3.4: Obtain the score function for triples through the similarity function.
7. The recommendation method for a learning knowledge graph that integrates interactive attention according to claim 1, characterized in that, In step 4, the multi-task learning sharing module trains the recommendation module through alternating training, and finally obtains the final predicted probability of user participation in the project through the prediction function, including: Step 4.1: For the user's initial feature vector, after feature embedding, the compressed features of the user's hidden layer are extracted through a multi-layer neural network. Step 4.2: For the original input vector of the item, the latent features of the item are calculated through the multi-task learning shared module; Step 4.3: Calculate the final predicted probability of user participation in the project using the prediction function.
8. A recommendation system that integrates interactive attention-based learning knowledge graphs, characterized in that, include: The data acquisition module is used to obtain user-item interaction information from the recommendation dataset and knowledge graph information from the knowledge graph dataset. Based on the historical records of user-rated items, a user-item interaction matrix is constructed. After establishing a mapping between entities in the knowledge graph dataset and the recommendation dataset, a knowledge graph is constructed. The user-item interaction information includes the set of users and items, and the historical records of user-rated items. The knowledge graph consists of head entity-relationship-tail entity triples. An interactive training preparation module is used to input item features and head entity features obtained from user-item interaction information into a multi-task learning and sharing module. The multi-task learning and sharing module comprises multiple feature attention cross-compression modules, each including a feature interaction attention unit and a cross-compression unit. Item features and head entity features are learned through the interaction attention unit to obtain item features and entity features fused with contextual information, preparing for interactive training of the recommendation module and knowledge graph module. This includes: Step 2.1: For each input candidate item and head entity vector, multiply it with the trainable parameter matrix to obtain the query vector and address vector of each vector; Step 2.2: Calculate the similarity score between the query vector and the address vector, and then apply a normalized exponential function to the similarity score to obtain their respective weight coefficient scores; Step 2.3: Perform a weighted summation of the weight coefficient scores with the item vector or the head entity vector to obtain the item self-attention score and the head entity self-attention score, respectively. Step 2.4: Perform a dot product between the item self-attention score and the head entity self-attention score and their self-attention scores processed by the activation function; Step 2.5: Add the dot product value to its initial feature vector to obtain the item vector and head entity vector respectively; Step 2.6: The item vector and the head entity vector are first cross-multiplied in the cross-compression layer to obtain the cross-multiplication matrix; Step 2.7: Transpose the cross-multiplication matrix to obtain the transposed cross-multiplication matrix; Step 2.8: Compress the cross-cutting features and calculate the feature vectors of the items and head entities in the next layer; The knowledge graph training module is used to train the knowledge graph module based on the multi-task learning sharing module through alternating training, and obtains the score function of triples through the similarity function; The recommendation module training module is used to train the recommendation module through alternating training based on the multi-task learning shared module, and finally obtains the final predicted probability of user participation in the project through the prediction function.