Semi-supervised learning-based interpretable recommendation method

Through an interpretable recommendation method based on semi-supervised learning, the label propagation law and collaborative filtering technology are used to process unlabeled data, combined with the knowledge graph, the high cost and 'cold start' problems in the recommendation system are solved, and the recommendation is realized System performance improvement and interpretability of recommended results.

CN118916545APending Publication Date: 2024-11-08GUANGZHOU PANYU POLYTECHNIC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410987961.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-23
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

The existing recommendation system is costly to obtain high-quality labeled data and faces the problem of ‘cold start’, resulting in poor user experience, especially when new users enter the system, they cannot push content of interest in time.

Method used

Using an interpretable recommendation method based on semi-supervised learning, unlabeled data is processed through label propagation laws and collaborative filtering technology, combining entity and relational attributes in the knowledge graph, metapaths and multi-hop connections are learned, and useful paths are generated dynamically, and through Graph convolutional network adjusts node embedding representations, generates Top-N project recommendations, and extracts critical paths between users and projects to provide a trusted explanation of recommended paths.

Benefits of technology

It effectively reduces the cost of obtaining high-quality labeled data, improves the performance and user experience of the recommendation system, provides interpretable recommendation results, and improves user stickiness.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118916545A_ABST
    Figure CN118916545A_ABST
Patent Text Reader

Abstract

The invention provides an interpretable recommendation method based on semi-supervised learning, and the method comprises the steps: processing the training data of labeled data and unlabeled data through a propagation rule of label propagation and a collaborative filtering technology, and combining the entity and relation attributes in a knowledge graph to improve the recommendation performance; learning meta-paths and multi-hop connections hidden in an original knowledge graph, dynamically generating useful meta-paths, and adjusting node embedding representation through a graph convolutional network; the recommendation interpretation layer aggregates different entity vectors, predicts the recommendation possibility of each item, and generates Top-N item recommendation; a critical path between a user and an item is extracted to provide a trusted recommended path interpretation. The problem that in the prior art, cost is high due to the fact that a large amount of high-quality annotation data needs to be obtained is solved. In addition, a critical path between the user and the item is extracted to provide a trusted recommended path interpretation. According to the invention, performance improvement of the recommendation system and interpretability of the recommendation result are realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of Internet technology, and in particular to an explainable recommendation method based on semi-supervised learning. Background Art

[0002] With the rapid development of technology today, the amount of data is also increasing day by day, and people are increasingly feeling helpless in the face of massive data. It is precisely to solve the problem of information overload that people have proposed recommendation systems. Recommendation systems are widely used in the fields of news, audio and video, aiming to provide users with content of interest and improve user experience. However, there are still some challenges in practical applications. For example, the cost of obtaining a large amount of high-quality labeled data is very high. Secondly, the recommendation system also faces the danger of "cold start", that is, when a new user enters the system, the content of interest to the user cannot be pushed in time, affecting the user experience.

[0003] Therefore, in order to improve user experience and increase user stickiness, it is important to solve the above problems. Summary of the invention

[0004] The purpose of the present invention is to provide an explainable recommendation method based on semi-supervised learning, which can directly process unlabeled data and solve the high cost problem caused by the need to obtain a large amount of high-quality labeled data in the prior art.

[0005] The objective of the present invention is achieved through the following technical solutions: The explainable recommendation method based on semi-supervised learning is characterized by comprising the following steps: Step S1: In the semi-supervised learning layer, the propagation law of label propagation and the collaborative filtering technology are used to process the training data of labeled data and unlabeled data, and the entity and relationship attributes in the knowledge graph are combined to improve the recommendation performance; Step S2: Learn the hidden meta-paths and multi-hop connections in the original knowledge graph in the KGTN layer, dynamically generate useful meta-paths, and adjust the node embedding representation through the graph convolutional network; Step S3: The recommendation explanation layer aggregates different entity vectors, predicts the likelihood of each item being recommended, and generates Top-N item recommendations; it extracts the key path between users and items to provide a credible recommendation path explanation.

[0006] Furthermore, the step S1 includes: Step S11: Construct a user-item interaction matrix using the interaction behaviors between users and items ; Step S12: Based on the user-item interaction matrix, users and items are combined into sequence pairs as nodes and input into the semi-supervised learning framework.

[0007] Furthermore, in the step S11, after constructing the user-item interaction matrix, the following steps are further included: Step S111: In the interaction matrix, a label value of 1 indicates that there is an interaction between user u and item v, and a label value of 0 indicates that there is no interaction between user u and item v. If there is no clear interaction information between user u and item v, the reserved label is represented as empty. Step S112: Use the graph neural network model to approximate the feature selection strategy. The model assigns weights to relevant feature information according to the attention mechanism. The higher the weight, the higher the relevance. Step S113: Set is the scalar value of the i-th feature matrix. Before training, the scalar values ​​of the feature matrix need to be set to equal values, and then linearly normalized using softmax. Through a fully connected layer, it is transformed into , each input feature is compared with the normalized Multiply, in The input matrix can be expressed as: ; in ,|| represents the concatenation operation. Through the above operations, the weight initialization value is 1 / L, which will change during the training process to indicate the importance of the feature.

[0008] Furthermore, the step S12 includes: Step S121: Add a self-loop to the interaction matrix to obtain a new adjacency matrix. , A and The diagonal matrices of ; Each node is associated with a d-dimensional feature vector, and the node feature matrix is ​​represented by X; Step S122: Input features Precompute and transform linear layers specific to each feature matrix.

[0009] Furthermore, the conversion of the linear layer specific to each feature matrix in step S122 includes: First, the output activation of the first layer is normalized using L2, and then the output features are connected using ReLU for nonlinear transformation, which is mapped to the second linear transformation layer. Finally, the output logits of the second layer are used to calculate the cross entropy loss, as shown in the following formula: ; in, , is the input feature, is the RELU activation function.

[0010] Furthermore, the step S2 comprises: Step S21: Select an adjacency matrix A from the candidate adjacency matrix set. 1 , and then use the convolution kernel and the weight of the softmax function to concatenate the adjacency matrix A 1 Calculate a new meta-path adjacency matrix , and then use another adjacency matrix A 2 Get another new meta-path adjacency matrix ; Step S22: Use matrix multiplication to obtain a new graph structure by combining the two new element path adjacency matrices It is expressed as: ; Step S23, the graph convolutional network is constructed by With feature representation A convolution operation is performed to update the embedding representation of the node, which is expressed as follows: ; in is the activation function, It is the first The feature representation of the layer, , yes The degree matrix of Is a trainable weight matrix; through training, we get The vector representation of each node below.

[0011] Furthermore, the step S3 includes: Step S31: Splice the hierarchical nodes into a whole node vector representation; for user node u and project node v, use the splicing operation || to splice the learning features in each layer into a vector , the formula is: ; Step S32: Calculate the click probability between the user and the item. The calculation formula is as follows: ; in, is the sigmoid function; Step S33: Use the new graph structure The critical path in As an explanation of the recommended items.

[0012] Furthermore, at least one of the following four indicators is used to evaluate the performance of the recommendation method: Hit rate: evaluates the accuracy of the model's recommendations, that is, whether the user's requirements are included in the model's recommendations; Normalized discounted cumulative gain: evaluates the position of the user's demand item in the model recommendation list, the higher the position, the better; Logarithmic loss: used to measure the difference between the model prediction results and the true labels; Accuracy: The ratio of the number of correct predictions in positive and negative examples to the total number.

[0013] The implementation case of the present invention uses the propagation law of label propagation and collaborative filtering and other technologies to process training data containing a small amount of labeled data and a large amount of unlabeled data in the semi-supervised learning layer, solving the high cost problem caused by the need to obtain a large amount of high-quality labeled data in the prior art. And combine the entity and relationship attributes in the knowledge graph to improve the recommendation performance; in the KGTN layer, learn the meta-paths and multi-hop connections hidden in the original knowledge graph, and adjust the node embedding representation; the recommendation explanation layer aggregates different entity vectors, predicts the possibility of each item being recommended, and generates Top-N item recommendations. In addition, the key path between users and items is extracted to provide a credible recommendation path explanation. The present invention achieves performance improvement of the recommendation system and the interpretability of the recommendation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A schematic diagram of the steps of the explainable recommendation method based on semi-supervised learning of the present invention; Figure 2 This is a recommendation model framework diagram of the explainable recommendation method based on semi-supervised learning of the present invention. DETAILED DESCRIPTION

[0015] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0016] The following describes the embodiments of the present disclosure through specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present disclosure.

[0017] The explainable recommendation method based on semi-supervised learning includes the following steps: Step S1: In the semi-supervised learning layer, the propagation law of label propagation and the collaborative filtering technology are used to process the training data containing a small amount of labeled data and a large amount of unlabeled data, and the entity and relationship attributes in the knowledge graph are combined to improve the recommendation performance.

[0018] Further, step S1 includes: Step S11: Construct a user-item interaction matrix using the interaction behaviors between users and items .

[0019] Furthermore, in step S11, after constructing the user-item interaction matrix, the following steps are further included: Step S111, in the matrix, a label value of 1 indicates that there is an interaction between user u and item v, and a label value of 0 indicates that there is no interaction between user u and item v. If there is no clear interaction information between user u and item v, the reserved label is represented as empty.

[0020] Step S112: Use the graph neural network model to approximate the feature selection strategy. The model assigns weights to relevant feature information according to the attention mechanism. The higher the weight, the higher the degree of relevance.

[0021] Step S113: Set is the scalar value of the i-th feature matrix. Before training, the scalar values ​​of the feature matrix need to be set to equal values, and then linearly normalized using softmax. Through a fully connected layer, it is transformed into , each input feature is compared with the normalized Multiply, in The input matrix can be expressed as: Formula (1) in ,|| represents the concatenation operation.

[0022] Through the above operations, the weights are initialized to 1 / L and will change during training to indicate the importance of the feature.

[0023] Step S12: Based on the user-item interaction matrix, users and items are combined into sequence pairs as nodes and input into the semi-supervised learning framework.

[0024] Further, step S12 includes: Step S121: Add a self-loop to the interaction matrix to obtain a new adjacency matrix. , A and The diagonal matrices of Each node is associated with a d-dimensional feature vector, and the node feature matrix is ​​represented by X.

[0025] Step S122: Input features Precompute and transform linear layers specific to each feature matrix.

[0026] Furthermore, in step S122, the conversion of the linear layer specific to each feature matrix includes: firstly, normalizing the output activation of the first layer using L2, then connecting the output features using ReLU for nonlinear transformation, thereby mapping to the second linear transformation layer, and finally calculating the cross entropy loss using the output logits of the second layer. As shown in the following formula: Formula (2) In formula (2), , is the input feature, is the RELU activation function.

[0027] Step S2: In the KGTN layer, the hidden meta-paths and multi-hop connections in the original knowledge graph are learned, useful meta-paths are dynamically generated, and the node embedding representation is adjusted through the graph convolutional network.

[0028] Further, step S2 includes: Step S21: First, select an adjacency matrix A from the candidate adjacency matrix set. 1 , and then use the convolution kernel and the weight of the softmax function to concatenate the adjacency matrix A 1 Calculate a new meta-path adjacency matrix , and then use another adjacency matrix A 2 Get another new meta-path adjacency matrix .

[0029] Step S22: Use matrix multiplication to obtain a new graph structure by combining the two new element path adjacency matrices It can be expressed as: Formula (3) Step S23, the graph convolutional network is constructed by With feature representation A convolution operation is performed to update the embedding representation of the node, which can be expressed by the following formula: Formula (4) In formula (4) is the activation function, It is the first The feature representation of the layer, , yes The degree matrix of is a trainable weight matrix. Through training, we get The vector representation of each node below.

[0030] Step S3: The recommendation explanation layer aggregates different entity vectors, predicts the likelihood of each item being recommended, and generates Top-N item recommendations. In addition, the key path between users and items is extracted to provide a credible recommendation path explanation.

[0031] Further, step S3 includes: Step S31: Splice the hierarchical nodes into a whole node vector representation. For user node u and project node v, use the splicing operation || to splice the learning features in each layer into a vector , the formula is: Formula (5) Step S32: Calculate the click probability between the user and the item. The calculation formula is as follows: Formula (6) In formula (6), is the sigmoid function.

[0032] Step S33: Use the new graph structure The critical path in As an explanation of the recommended items.

[0033] The interpretable recommendation method based on semi-supervised learning of the present invention aims to improve the performance of the recommendation system and the interpretability of the recommendation results, and uses the following four indicators to evaluate the performance of the recommendation method: A. HR (hit rate): emphasizes the accuracy of model recommendations, that is, whether the user's requirements are included in the model's recommendations; B. NDCG (Normalized Discounted Cumulative Gain): emphasizes the position of the user's demand item in the model recommendation list, the higher the better; C.LogLoss (logarithmic loss): used to measure the difference between the model prediction results and the true label; D.ACC (accuracy): The ratio of the number of correct predictions in positive and negative examples to the total number.

[0034] The experimental results of the explainable recommendation method of the present invention are compared with those of RippleNet, KGCN, KGAT, KGNN-LS, MKR and KGFER models on the BookCrossing dataset, as shown in Table 1-2. As can be seen from Table 1-2, the proposed algorithm can obtain the best results in most evaluation indicators.

[0035] Table 1 Comparison of hit rates of different algorithms Table 2 Comparison of normalized discounted cumulative gains of different algorithms The experimental results of the explainable recommendation method of the present invention are compared with those of LPA, GCN, GAT, LPA-GCN and mma models on the BookCrossing dataset, as shown in Table 3. As can be seen from Table 3, the proposed algorithm can obtain the best results on most evaluation indicators.

[0036] Table 3 Comparison of logarithmic loss and accuracy of different algorithms In summary, the present invention can utilize collaborative knowledge graphs and semi-supervised learning to achieve performance improvement and recommendation results of the recommendation system, and performs well on the data set.

[0037] The above is only for explaining the implementation mode of the present invention and is not intended to limit the present invention. For those skilled in the art, any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention without creative work should be included in the protection scope of the present invention.

Claims

1. An explainable recommendation method based on semi-supervised learning, characterized in that: The following steps are involved: Step S1: In the semi-supervised learning layer, the propagation law of label propagation and the collaborative filtering technology are used to process the training data of labeled data and unlabeled data, and the entity and relationship attributes in the knowledge graph are combined to improve the recommendation performance; Step S2: Learn the hidden meta-paths and multi-hop connections in the original knowledge graph in the KGTN layer, dynamically generate useful meta-paths, and adjust the node embedding representation through the graph convolutional network; Step S3: The recommendation explanation layer aggregates different entity vectors, predicts the likelihood of each item being recommended, and generates Top-N item recommendations; it extracts the key path between users and items to provide a credible recommendation path explanation.

2. The explainable recommendation method based on semi-supervised learning according to claim 1, characterized in that: The step S1 comprises: Step S11: Construct a user-item interaction matrix using the interaction behaviors between users and items ; Step S12: Based on the user-item interaction matrix, users and items are combined into sequence pairs as nodes and input into the semi-supervised learning framework.

3. The explainable recommendation method based on semi-supervised learning according to claim 2 is characterized in that In the step S11, after constructing the user-item interaction matrix, the following steps are further included: Step S111: In the interaction matrix, a label value of 1 indicates that there is an interaction between user u and item v, and a label value of 0 indicates that there is no interaction between user u and item v. If there is no clear interaction information between user u and item v, the reserved label is represented as empty. Step S112: Use the graph neural network model to approximate the feature selection strategy. The model assigns weights to relevant feature information according to the attention mechanism. The higher the weight, the higher the relevance. Step S113: Set is the scalar value of the i-th feature matrix. Before training, the scalar values ​​of the feature matrix need to be set to equal values, and then linearly normalized using softmax. Through a fully connected layer, it is transformed into , each input feature is compared with the normalized Multiply, in The input matrix can be expressed as: ; in ,|| represents the concatenation operation. Through the above operations, the weight initialization value is 1 / L, which will change during the training process to indicate the importance of the feature.

4. The explainable recommendation method based on semi-supervised learning according to claim 2, characterized in that: The step S12 comprises: Step S121: Add a self-loop to the interaction matrix to obtain a new adjacency matrix. , A and The diagonal matrices of ; Each node is associated with a d-dimensional feature vector, and the node feature matrix is ​​represented by X; Step S122: Input features Precompute and transform linear layers specific to each feature matrix.

5. The explainable recommendation method based on semi-supervised learning according to claim 4, characterized in that: The conversion of the linear layer specific to each feature matrix in step S122 includes: First, the output activation of the first layer is normalized using L2, and then the output features are connected using ReLU for nonlinear transformation, which is mapped to the second linear transformation layer. Finally, the output logits of the second layer are used to calculate the cross entropy loss, as shown in the following formula: ; in, , is the input feature, is the RELU activation function.

6. The explainable recommendation method based on semi-supervised learning according to claim 1, characterized in that: The step S2 comprises: Step S21: select adjacency matrix A1 from the candidate adjacency matrix set, and then use the convolution kernel and the weight of the softmax function to calculate the adjacency matrix A1 to obtain a new meta-path adjacency matrix , and then use another adjacency matrix A2 to get another new meta-path adjacency matrix ; Step S22: Use matrix multiplication to obtain a new graph structure by combining the two new element path adjacency matrices It is expressed as: ; Step S23, the graph convolutional network is constructed by With feature representation A convolution operation is performed to update the embedding representation of the node, which is expressed as follows: ; in is the activation function, It is the first The feature representation of the layer, , yes The degree matrix of Is a trainable weight matrix; through training, we get The vector representation of each node below.

7. The explainable recommendation method based on semi-supervised learning according to claim 1, characterized in that: The step S3 includes: Step S31: Splice the hierarchical nodes into a whole node vector representation; for user node u and project node v, use the splicing operation || to splice the learning features in each layer into a vector , the formula is: ; Step S32: Calculate the click probability between the user and the item. The calculation formula is as follows: ; in, is the sigmoid function; Step S33: Use the new graph structure The critical path in As an explanation of the recommended items.

8. The explainable recommendation method based on semi-supervised learning according to any one of claims 1 to 7, characterized in that: The performance of the recommended method is evaluated using at least one of the following four indicators: Hit rate: evaluates the accuracy of the model's recommendations, that is, whether the user's requirements are included in the model's recommendations; Normalized discounted cumulative gain: evaluates the position of the user's demand item in the model recommendation list, the higher the position, the better; Logarithmic loss: used to measure the difference between the model prediction results and the true labels; Accuracy: The ratio of the number of correct predictions in positive and negative examples to the total number.