Contrast learning-based enhanced graph convolutional network recommendation method applied to recommendation system

Abstract

The invention discloses an enhanced graph convolutional network recommendation method based on comparative learning and applied to a recommendation system, and belongs to the technical field of personalized recommendation. The method comprises the following steps: firstly, constructing a user-project interaction matrix by analyzing interaction behaviors of a user and a project, additionally enhancing an initial interaction matrix, and generating a contrast matrix with the initial interaction matrix; respectively inputting the two groups of user-item contrast matrixes into a multilayer graph convolutional network to generate two groups of different enhanced user-item correlation matrixes; through recursive propagation of multilayer graph convolution, deep characteristics of interaction information can be captured, and richer embedded representations are generated for users and projects. Fusing the two groups of embedding matrixes for comparison to generate final user embedding and project embedding; and finally generating a recommendation result by using a collaborative filtering algorithm. According to the method, through message passing of comparative learning and multilayer graph convolution, implicit association of a user-item is drawn close, and a better recommendation effect can be achieved when recommendation is performed to the user.

Description

Technical Field

[0001] The present invention belongs to the technical field of personalized recommendation, and specifically relates to an enhanced graph convolutional network recommendation method based on contrastive learning applied to a recommendation system. Background Art

[0002] Personalized recommendation has been applied to many online services, such as e-commerce, advertising, and social media. Learning vector representation is the core of modern recommendation methods, from early matrix decomposition to the deep learning-based methods that have emerged in recent years, all of which are intelligent algorithms proposed based on personalized recommendation.

[0003] Existing personalized recommendation methods obtain the embedding of users (or items) by mapping from pre-existing features that describe users (or items). Some methods extend the embedding function of matrix factorization by integrating deep representations learned from rich side information of items, or use neural collaborative filtering models to replace the interaction function of matrix factorization with nonlinear neural networks. Projection-based collaborative filtering models use Euclidean distance metrics as interaction functions, etc. Despite their effectiveness, these methods are insufficient to generate satisfactory embedding collaboration signals. Specifically, most existing methods only use descriptive features without considering user-item interactions.

[0004] To address the problem of insufficient interactive information capture, the present invention uses a graph convolutional network (GCN) in the selection of the recommendation framework. In view of the problem that recommendation methods often face sparse user-item interaction data, the idea of ​​contrastive learning is introduced. Contrastive learning can mine more potential information from limited data by constructing positive and negative samples, thereby improving the model's generalization ability for sparse data. Summary of the invention

[0005] In order to solve the problem that the current personalized recommendation methods do not capture enough interactive signals of potential users (or projects), which are not encoded in the embedding process, so when recommending to a certain user, the best project is often not recommended. This paper proposes an enhanced graph convolutional network recommendation method based on contrastive learning for recommendation systems. Through contrastive learning and multi-layer graph convolution message passing, the implicit association between users and projects is shortened, and better recommendation effects can be achieved when recommending to users.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is:

[0007] An enhanced graph convolutional network recommendation method based on contrastive learning applied to a recommendation system comprises the following steps:

[0008] Step 1: Extract the initial interaction matrix related to user-item based on the historical interaction behavior data between user and item; additionally enhance the initial interaction matrix, generate a comparison matrix with the initial interaction matrix, and initialize the weight matrix of user and item;

[0009] Step 2: The comparison matrix and weight matrix are recursively propagated through multi-layer graph convolution to obtain the deep characteristics of the interactive information and generate richer embedding representations for users and items. Combined with the idea of ​​contrastive learning, the two sets of embedding matrices for contrast are fused to generate the final user embedding and item embedding.

[0010] Step 3: Calculate the user's preference for the project through the inner product of user embedding and project embedding, and generate recommendation results in combination with collaborative filtering algorithm.

[0011] As a preferred technical solution of the present invention, the step 1 specifically includes:

[0012] (1) Initialize the user-item weight matrix W using a normal distribution with a standard deviation of 0.1. The formula is:

[0013] W~N(0,0.1 2 )(1)

[0014] (2) Get an initial sparse graph A, where each user and item is a vertex. A is represented as:

[0015]

[0016] In formula (2), R represents the interaction matrix between users and items;

[0017] (3) Final sparse graph It is expressed as:

[0018]

[0019] In formula (3), D is the degree matrix;

[0020] Initial Embedding and Respectively expressed as:

[0021]

[0022] As a preferred technical solution of the present invention, the step 2 specifically includes:

[0023] (1) The recursive propagation process of multi-layer graph convolution is equivalent to:

[0024]

[0025] In formulas (6) and (7), k is the number of times the graph convolution is performed;

[0026] (2) Using the idea of ​​contrastive learning, we obtain two sets of embedding matrices E1 and E2 after multi-layer graph convolution recursive propagation, which are expressed as:

[0027]

[0028] (3) Aggregate the two sets of contrasting embedding matrices to obtain the final embedding matrix for model prediction:

[0029] E=(E1+E2) / 2(10).

[0030] As a preferred technical solution of the present invention, the step three specifically includes:

[0031] (1) Perform inner product to estimate the user's preference for the target item:

[0032]

[0033] In formula (11), e u and e i are user embedding and item embedding, obtained from the final embedding matrix E;

[0034] (2) Obtain the user's ratings for all items and take the top N largest ratings to generate the Top-N recommendation results.

[0035] In addition, the present invention also proposes a recommendation system for implementing the above recommendation method, including a data initialization module, a message transmission module, a comparison aggregation module and a prediction recommendation module, wherein:

[0036] The data initialization module implements:

[0037] Obtain user-item historical interaction behavior data and initialize weights;

[0038] The messaging module implements:

[0039] The initialized weight matrix and the constructed user-item interaction matrix are passed through multiple layers of graph convolution to generate a deeper embedding matrix.

[0040] The comparison aggregation module implements:

[0041] Aggregate the constructed comparison data to generate the final interaction information;

[0042] The prediction and recommendation module implements:

[0043] The final interaction information is extracted to generate deep embedding representations of users and items respectively; inner product operations are performed through the embedding representations of users and items to predict the user's preference score for the item.

[0044] The present invention proposes an enhanced graph convolutional network recommendation method based on contrastive learning for use in recommendation systems, which can recommend more suitable items to users. The recommendation method first extracts the user-item correlation matrix based on the data of historical user-item interaction behaviors; secondly, the initialization weight matrix is ​​added to the recursive propagation of multi-layer graph convolution to capture deeper user-item embedding information; then the embedding information of the control group is added using the idea of ​​contrastive learning, and the two sets of embedding information are merged into the final user-item embedding; finally, the user's preference for the item is predicted by the inner product of the user embedding and the item embedding, and the recommendation result is generated using the collaborative filtering algorithm. Compared with the prior art, the beneficial effects of the present invention are:

[0045] (1) Before the recommendation algorithm runs, the present invention preprocesses the data, including randomly initializing the weight matrices of users and items. These weight matrices are also recursively propagated through multi-layer graph convolutions, thereby improving the propagation effect of feature information and the ability to capture interactive features. Finally, by calculating the inner product of user embedding and item embedding, the user's preference for the item is evaluated and the recommendation result is generated.

[0046] (2) The performance of the present invention on a variety of data sets is superior. The most advanced recommendation method is used for comparative experiments. In order to reflect the fairness of the experiment, no parameters of the method are changed to ensure the best performance. Finally, all evaluation indicators on most data sets have advantages. Therefore, the present invention can effectively propagate feature information, capture the deep interaction characteristics of users and items, and learn more expressive embedding representations by designing the initialization strategy of the weight matrix and the multi-layer graph convolutional network, thereby significantly improving the expressiveness of the model and the recommendation effect.

[0047] (3) This paper solves the problem of personalized recommendation by using the idea of ​​contrastive learning and using multi-layer GCN to capture deeper user-item interaction information. This is a natural way to encode collaborative signals in the interaction graph structure, which improves the scalability and usability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a structural framework diagram of the model constructed for the present invention.

[0049] Figure 2 An example graph is constructed for user recommendation, where a represents the user's feature relationship and b represents the constructed simplified graph structure. DETAILED DESCRIPTION

[0050] Example 1

[0051] See also Figure 1 As shown, this embodiment proposes an enhanced graph convolutional network recommendation method based on contrastive learning applied to a recommendation system, comprising the following steps:

[0052] Step 1: Extract the initial interaction matrix related to user-item based on the historical interaction behavior data between user and item; additionally enhance the initial interaction matrix, generate a comparison matrix with the initial interaction matrix, and initialize the weight matrix of user and item. Specifically include:

[0053] (1) Initialize the user-item weight matrix W using a normal distribution with a standard deviation of 0.1. The formula is:

[0054] W~N(0,0.1 2 )(1)

[0055] (2) Get an initial sparse graph A, where each user and item is a vertex. A is represented as:

[0056]

[0057] In formula (2), R represents the interaction matrix between users and items;

[0058] (3) Final sparse graph It is expressed as:

[0059]

[0060] In formula (3), D is the degree matrix;

[0061] Initial Embedding and Respectively expressed as:

[0062]

[0063] Step 2: Recursively propagate the above-mentioned comparison matrix and weight matrix through multi-layer graph convolution to obtain the deep characteristics of the interactive information and generate richer embedding representations for users and items; then combine the idea of ​​contrastive learning to fuse the two sets of embedding matrices for contrast to generate the final user embedding and item embedding. Specifically include:

[0064] (1) The recursive propagation process of multi-layer graph convolution is equivalent to:

[0065]

[0066] In formulas (6) and (7), k is the number of times the graph convolution is performed;

[0067] (2) Using the idea of ​​contrastive learning, we obtain two sets of embedding matrices E1 and E2 after multi-layer graph convolution recursive propagation, which are expressed as:

[0068]

[0069] (3) Aggregate the two sets of contrasting embedding matrices to obtain the final embedding matrix for model prediction:

[0070] E=(E1+E2) / 2(10)

[0071] Step 3: Calculate the user's preference for the item through the inner product of the user embedding and the item embedding, and generate the recommendation results by combining the collaborative filtering algorithm. Specifically include:

[0072] (1) Perform inner product to estimate the user's preference for the target item:

[0073]

[0074] In formula (11), e u and e i are user embedding and item embedding, obtained from the final embedding matrix E;

[0075] (2) Obtain the user's ratings for all items and take the top N largest ratings to generate the Top-N recommendation results.

[0076] Example 2

[0077] This embodiment proposes a recommendation system for implementing the above recommendation method, including a data initialization module, a message transmission module, a comparison aggregation module and a prediction recommendation module, wherein:

[0078] The data initialization module implements:

[0079] Obtain user-item historical interaction behavior data and initialize weights;

[0080] The messaging module implements:

[0081] The initialized weight matrix and the constructed user-item interaction matrix are passed through multiple layers of graph convolution to generate a deeper embedding matrix.

[0082] The comparison aggregation module implements:

[0083] Aggregate the constructed comparison data to generate the final interaction information;

[0084] The prediction and recommendation module implements:

[0085] The final interaction information is extracted to generate deep embedding representations of users and items respectively; inner product operations are performed through the embedding representations of users and items to predict the user's preference score for the item.

[0086] Example 3

[0087] Taking music recommendation as an example, the recommendation system proposed by the present invention is introduced:

[0088] The recommendation system infers user preferences from user-item interactions or static features, and further recommends items that the user may be interested in. The items here can be some of the user's interests and hobbies, such as music recommendations, food recommendations, and online shopping product recommendations.

[0089] Take music as an example, assume that user 1 likes music 1, music 2, and music 3; user 2 likes music 2, music 4, and music 5; user 3 likes music 3 and music 4. In this embodiment, users 1, 2, and 3 are represented by u1, u2, and u3, respectively, and music 1-5 are represented by i1, i2, ..., i5, respectively. Figure 2 shown.

[0090] Figure 2 Part a represents the user's feature relationship, and part b is a simple graph structure constructed for recommendation analysis. Figure 2 In (b), the path u1←i2←u2 indicates the behavioral similarity between u1 and u2, because both users interacted with i2; the longer path u1←i2←u2←i4 indicates that u1 is likely to adopt i4, because u1's similar user u2 has previously used i4. In addition, from an overall perspective, music i4 is more likely to arouse u1's interest than music i5, because there are two paths connecting<i4,u1> , and there is only one path connecting<i5,u1> Therefore, recommending item i4 to user u1, that is, recommending the song music 4 to user 1, can achieve the best recommendation result.

[0091] Example 4 Comparative Experiment

[0092] 1. Dataset Introduction

[0093] The datasets used are Gowalla, douban, Amazon-Book, Yelp2018, coat, yahoo, Food, and ml-1m. The datasets are from:

[0094] Gowalla: Check-in dataset obtained from Gowalla.

[0095] Douban: From Douban, a well-known Chinese review website.

[0096] Amazon-Book: The Amazon Book dataset is a comprehensive data set that focuses on book reviews on the Amazon platform.

[0097] Yelp2018: Business reviews and user data from Yelp from the 2018 edition of the Yelp Challenge.

[0098] ml-1m: A widely used dataset collected from Movielens. Our experiments use the 1M version which converts explicit data into implicit feedback and treats all user-item ratings as positive interactions.

[0099] Food: The Food dataset contains recipe details and reviews from Food.com (formerly GeniusKitchen). The data includes cooking recipes and review text.

[0100] Coat, Yahoo: The Coat and Yahoo datasets are from clothing shopping recommendation services and Yahoo Music, respectively. Both datasets contain a training set of biased rating data collected from normal user interactions and a test set of unbiased rating data containing user ratings on randomly selected items.

[0101] The statistics of these datasets are shown in Table 1.

[0102] Table 1 Statistics of the dataset

[0103]

[0104] 2. Evaluation indicators

[0105] In the recommendation system, the main function of evaluation indicators is to quantify the performance of the recommendation system, help developers understand the pros and cons of the model, and guide improvement and optimization. The evaluation indicators used in this experiment.

[0106] NDCG (Normalized Discounted Cumulative Gain) is a ranking-based indicator that emphasizes whether the order of related items in the recommendation list is reasonable. It calculates the benefits of related items in the recommendation results through the concept of cumulative gain (Cumulative Gain, CG), and introduces a discount factor to reduce the contribution value of related items with lower rankings. The result is then normalized to between 0 and 1. Its function is to measure the relevance and ranking quality of the results in the recommendation list. NDCG is an important indicator when it is necessary to emphasize the ranking accuracy of the recommendation results (such as movie and music recommendations).

[0107] Precision: Precision indicates the proportion of relevant items in the recommended list, that is, the correctness of the recommendation. It measures the proportion of relevant items in the recommendation results. When the size of the recommendation list is fixed and more attention is paid to the quality of the recommendation rather than the coverage, Precision is the main indicator (such as product recommendations in e-commerce).

[0108] Recall, Recall represents the proportion of all relevant items found by the recommendation system, that is, the coverage of the recommendation. It measures the proportion of relevant items that the user may be interested in that are captured by the recommendation system. When the recommendation system covers the content that the user is interested in as much as possible, Recall is a key indicator.

[0109] 3 Experimental results

[0110] In order to evaluate the performance of the recommendation system proposed in the present invention, the experimental results of DR-GNN are compared. DR-GNN (a recommendation system based on distributed blue stick graph) is from: https: / / github.com / . Experiments are carried out according to the basic settings of DR-GNN to give full play to the best performance of DR-GNN. The experimental results of the present invention and DR-GNN on 8 data sets are shown in Table 2, where the bold part is the best part of the comparison data.

[0111] Table 2 Experimental results

[0112]

[0113]

[0114] By comparing the experimental results, it can be found that the present invention has achieved good recommendation effects on many data sets. This is due to the fact that the user-item interaction information is passed through contrastive learning and multi-layer graph convolution, which brings the implicit association between users and items closer, and can achieve better recommendation effects when recommending to users. Therefore, the present invention can capture deeper user-item interaction information in the data of the recommendation system, which can better help with recommendations.

[0115] The above contents are merely examples and explanations of the concept of the present invention. The technicians in this technical field may make various modifications or additions to the specific embodiments described or replace them in a similar manner. As long as they do not deviate from the concept of the invention or exceed the scope defined by the claims, they should all fall within the protection scope of the present invention.

Claims

1. An enhanced graph convolutional network recommendation method based on contrastive learning applied to recommendation systems, characterized in that: The steps include: Step 1: Extract the initial interaction matrix related to user-item based on the historical interaction behavior data between user and item; additionally enhance the initial interaction matrix, generate a comparison matrix with the initial interaction matrix, and initialize the weight matrix of user and item; Step 2: Recursively propagate the above-mentioned comparison matrix and weight matrix through multi-layer graph convolution to obtain the deep characteristics of the interactive information and generate richer embedding representations for users and items; Combined with the idea of ​​contrastive learning, the two sets of embedding matrices for comparison are fused to generate the final user embedding and item embedding; Step 3: Calculate the user's preference for the project through the inner product of user embedding and project embedding, and generate recommendation results in combination with collaborative filtering algorithm.

2. The recommendation method according to claim 1, characterized in that: The step 1 specifically includes: (1) Initialize the user-item weight matrix W using a normal distribution with a standard deviation of 0.

1. The formula is: W~N(0,0.1 2 )(1) (2) Get an initial sparse graph A, where each user and item is a vertex. A is represented as: In formula (2), R represents the interaction matrix between users and items; (3) Final sparse graph It is expressed as: In formula (3), D is the degree matrix; Initial Embedding and Respectively expressed as:

3. The recommendation method according to claim 2, characterized in that: The step 2 specifically includes: (1) The recursive propagation process of multi-layer graph convolution is equivalent to: In formulas (6) and (7), k is the number of times the graph convolution is performed; (2) Using the idea of ​​contrastive learning, we obtain two sets of embedding matrices E1 and E2 after multi-layer graph convolution recursive propagation, which are expressed as: (3) Aggregate the two sets of contrasting embedding matrices to obtain the final embedding matrix for model prediction: E=(E1+E2) / 2(10).

4. The recommendation method according to claim 3, characterized in that: The step three specifically includes: (1) Perform inner product to estimate the user's preference for the target item: In formula (11), e u and e i are user embedding and item embedding, obtained from the final embedding matrix E; (2) Obtain the user's ratings for all items and take the top N largest ratings to generate the Top-N recommendation results.

5. A recommendation system for implementing the recommendation method according to any one of claims 1 to 4, characterized in that: It includes data initialization module, message passing module, comparison aggregation module and prediction and recommendation module, among which: The data initialization module implements: Obtain user-item historical interaction behavior data and initialize weights; The messaging module implements: The initialized weight matrix and the constructed user-item interaction matrix are passed through multiple layers of graph convolution to generate a deeper embedding matrix. The comparison aggregation module implements: Aggregate the constructed comparison data to generate the final interaction information; The prediction and recommendation module implements: The final interaction information is extracted to generate deep embedding representations of users and items respectively; inner product operations are performed through the embedding representations of users and items to predict the user's preference score for the item.

Images