Cross-domain commodity recommendation method based on multilayer convolution blocks and comparative learning

Abstract

The invention belongs to the technical field of recommendation methods, and particularly relates to a cross-domain commodity recommendation method based on multilayer convolution blocks and comparative learning. And the multi-layer convolution block can effectively extract high-dimensional features of the commodity, including multi-modal information such as images and texts, so that the accuracy of a recommendation system is improved. And in comparison learning, feature representation learning can be improved by constructing positive and negative sample pairs, and the generalization ability of the model on cross-domain commodity recommendation is enhanced, so that the model has a better migration effect among different fields. The cross-domain recommendation can utilize commodity data of multiple fields to carry out joint training, the problem of data sparsity of a certain field can be relieved, and therefore the coverage rate of recommendation is increased. According to the method, the characteristic relation of cross-domain similar commodities can be effectively mined through a comparative learning mechanism, and the cold start problem of a recommendation system in a new commodity or new user scene is relieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of recommendation methods, and in particular relates to a cross-domain product recommendation method based on multi-layer convolutional blocks and contrastive learning. Background Art

[0002] With the rapid development of the Internet and the explosive growth of information, users are facing an increasingly prominent problem of information overload. The choices are becoming more complex, and it is becoming increasingly difficult to find products that meet their needs. In order to meet this challenge, recommendation systems have become a hot research topic that has attracted much attention. Through personalized recommendations, recommendation systems can help users filter out the most relevant content from massive amounts of information. As an information filtering tool, recommendation systems predict users' interest preferences and recommend products that they may be interested in, thereby improving users' decision-making efficiency.

[0003] At present, non-cross-domain recommendation systems mainly rely on user behavior data on a single platform or field for modeling. This approach has several significant problems. First, because user behavior data is limited to a single field, the data sparsity problem is exacerbated, especially for new users or long-tail products. The recommendation effect is poor. Secondly, there are differences in user preferences between different fields or platforms. It is difficult for a single-domain recommendation system to effectively capture the behavioral characteristics of users on other platforms, which limits the generalization ability of the model. Finally, when faced with diversified user needs, a single-domain recommendation system is difficult to provide comprehensive recommendations across platforms and fields, and cannot meet the complex needs of users.

[0004] Cross-domain product recommendation is a hot research direction in the current recommendation system field, aiming to solve the recommendation problem in different platforms, fields or scenarios. Traditional recommendation systems usually rely on user behavior data on a single platform or field for modeling, but in real life, users' purchasing behavior often involves multiple platforms and fields, which leads to data sparsity and field offset problems, affecting the recommendation effect. In order to solve these challenges, in recent years, cross-domain recommendation methods based on deep learning have gradually become a research hotspot, especially the model combining multi-layer convolutional neural network and contrastive learning, which can better extract cross-domain features of products and reduce the impact of field offset. Although cross-domain recommendation systems have made some progress in solving data sparsity and improving recommendation generalization capabilities, there are still many challenges and problems. 1. Cross-domain recommendation systems face the problem of domain differences. There are large differences in user behavior and product characteristics between different fields. How to effectively capture and integrate these heterogeneous data is still a difficulty. 2. Cross-domain recommendation is easily restricted by data privacy and security when sharing data. Data between different platforms is difficult to flow freely, which has an impact on model training and application. 3. The problem of feature alignment between domains has also attracted much attention. How to find consistent and relevant features in different domains and establish efficient cross-domain mapping relationships is still a key problem that needs to be solved. 4. The existing cross-domain recommendation methods have high computational complexity when processing large-scale data and are difficult to adapt to real-time and high-efficiency requirements. Summary of the invention

[0005] The purpose of the present invention is to solve the above problems. A cross-domain product recommendation method based on multi-layer convolutional blocks and contrastive learning is proposed to solve the data sparsity and cold start problems in cross-domain recommendation. The deep features of the products are extracted through multi-layer convolutional blocks, and a more effective association is established between different domains in combination with the contrastive learning mechanism, thereby improving the generalization ability and robustness of the recommendation system. This method effectively optimizes the overall performance of the recommendation system and improves the accuracy and usability of the recommendation.

[0006] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solutions:

[0007] A cross-domain product recommendation method based on multi-layer convolutional blocks and contrastive learning includes the following steps:

[0008] S1. Data preprocessing: We use the data from three different fields, "Books", "CDs_and_Vinyl" and "Movies_and_TV", in Amazon's user behavior data as the original data set and preprocess the data.

[0009] S2, model training: First, construct the source domain sequence graph, the target domain sequence graph and the cross-domain global graph based on the user behavior data. Second, pass the three sequence graphs through multi-layer encoders to obtain the corresponding item embeddings. Finally, the nonlinear layer and contrastive learning module are used to obtain the final intra-domain item representation embedding and inter-domain user preference embedding.

[0010] S3, nonlinear layer and contrastive loss: First, nonlinear projection is performed on the obtained embedding, and the SoftMax function is used to make the project embedding obtain richer information; in the contrastive learning module, the InfoNCE function is used to calculate the loss, maximize the mutual information between positive sample pairs, and minimize the consistency between negative sample pairs;

[0011] S4, prediction results: After obtaining the final intra-domain item representation view and inter-domain user preference view, the final prediction results are obtained through the contrast learning module and the link operation;

[0012] S5. Model evaluation: The model performance is evaluated and compared with other models by calculating the recall rate Recall and the normalized discounted cumulative gain NDCG.

[0013] Further, as a preferred technical solution of the present invention, step S1 includes the following steps:

[0014] S101, read the user's behavior data from the data sets in three fields, namely user ID, item ID, rating and timestamp, and delete repeated User-Item pairs;

[0015] S102. To alleviate the cold start problem, discard the records with less than 5 user or item interactions, and merge the three types of user data to reduce data sparsity and improve the generalization ability of the model.

[0016] S103. During the sampling process, a popularity calculation method is used to randomly sample negative samples for each user, and finally the training set, validation set, and test set are divided according to a ratio of 7:2:1.

[0017] Further as a preferred technical solution of the present invention, step S2 comprises the following steps:

[0018] S201, encoding the user's historical behavior data to obtain an input sequence, and obtaining a cross-domain global graph, a source domain sequence graph, and a target domain sequence graph from the input sequence;

[0019] S202: Input the three sequence graphs into the corresponding multi-layer encoder, and after multiple layers of convolution, obtain the source domain sequence embedding E S , target domain sequence embedding E T and global sequence embedding E;

[0020] S203. In order to alleviate the problem of data sparsity in cross-domain recommendation, the obtained embeddings are subjected to comparative learning respectively, so that user preferences in different fields complement each other and learn self-supervision signals to maximize the mutual information between them.

[0021] Further as a preferred technical solution of the present invention, step S3 comprises the following steps:

[0022] S301. In the prediction module, in order to obtain nonlinear information, nonlinear projection is applied to the output of the multi-layer encoder layer of each module, and the ReLU activation function is used to make the embedding have nonlinear characteristics, thereby improving the representation ability of the model;

[0023] S302. In the contrastive learning module, InfoNCE is used as the loss function to maximize the similarity between the positive sample and the target and minimize the similarity between the negative sample and the target, so as to help the model learn effective feature representation and back-propagate the learned parameters.

[0024] Further as a preferred technical solution of the present invention, step S4 comprises the following steps:

[0025] S401, finally get the source domain sequence embedded in E S and the target domain sequence embedding E T Afterwards, due to the inconsistency of project features in different fields, the data in the target domain will be relatively sparse, and a link operation is used to aggregate the embedding representations of the source and target domains to fully utilize user information in different fields;

[0026] S402. In the final prediction module, the binary cross entropy function is used as the loss function of the model. The loss is calculated by measuring the difference between the predicted probability distribution and the true label, and the model parameters are updated to enable the model to output more accurate prediction results.

[0027] Further, as a preferred technical solution of the present invention, multi-layer convolution blocks use lightweight convolution modules for multi-layer stacking, and the propagation rule between each layer of convolution blocks is:

[0028]

[0029] in, is the embedding representation of each layer of users, N u is the degree of the corresponding layer User, N i is the degree of the corresponding layer Item, is a regularization term. For each layer of embedding matrix representation, there are:

[0030]

[0031] Among them, the adjacency matrix A represents the number of non-zero entries in the i-th row vector, is a symmetric normalization term.

[0032] Further, as a preferred technical solution of the present invention, in order to make the obtained embedded representation have nonlinear characteristics, nonlinear projection is applied at the exit of each multi-layer convolution block:

[0033] h=RELU(W*Embedding(x)+b) (3)

[0034] Among them, W is the weight matrix of the activation function, Embedding(x) is the embedding vector, b is the bias, and RELU is the activation function to improve the representation ability of the model;

[0035] For the contrastive learning module, the InfoNCE function is used as the loss function:

[0036]

[0037] Among them, x i is the embedding of the sample, x j is the positive sample embedding corresponding to the sample, x k is the negative sample embedding corresponding to the sample, sim(x i ,x j ) is the cosine similarity between samples, τ is the smoothing coefficient;

[0038] Therefore, for the embedded preference learning, we have:

[0039]

[0040] Further, as a preferred technical solution of the present invention, due to the differences in the characteristics of projects in different fields, the data in the target domain is usually sparse; in order to make full use of user information in different fields, the E S and E L For aggregation, there are:

[0041]

[0042] For the cross-domain recommendation model, the binary cross entropy function is used as the loss function:

[0043]

[0044] Among them, j is the negative sampling term, σ(.) is the sigmoid function, δ(x t+1 ) is a filler item.

[0045] Further, as a preferred technical solution of the present invention, the recall rate Recall is calculated to represent the proportion of all truly relevant samples that are successfully retrieved by the model:

[0046]

[0047] Among them, TP is the number of samples correctly predicted by the model as positive, and FN is the number of samples that are actually positive but failed to be identified by the model;

[0048] Calculate the normalized discounted cumulative gain NDCG to measure the quality of the ranking results of the cross-domain recommendation model:

[0049]

[0050] Among them, rel is the relevance score of the i-th sample, is the relevance score of the result at the ith position in the ideal sorting case, and k is the first k results considered when calculating NDCG.

[0051] The cross-domain product recommendation method based on multi-layer convolutional blocks and contrastive learning described in the present invention has the following technical effects compared with the prior art by using the above technical solution:

[0052] (1) Improve feature extraction capabilities: Multi-layer convolutional blocks can effectively extract high-dimensional features of products, including multimodal information such as images and text, thereby improving the accuracy of the recommendation system.

[0053] (2) Enhance the generalization ability of the model: Contrastive learning can improve feature representation learning by constructing positive and negative sample pairs, thereby enhancing the generalization ability of the model for cross-domain product recommendations and achieving better migration effects between different domains.

[0054] (3) Alleviating the problem of data sparsity: Cross-domain recommendations can utilize product data from multiple fields for joint training, which helps alleviate the problem of data sparsity in a certain field and thus improves the coverage of recommendations.

[0055] (4) Reduce the impact of cold start: This method can effectively mine the feature relationships of similar products across domains through a comparative learning mechanism, thereby alleviating the cold start problem of the recommendation system in new product or new user scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 The figure is a flow chart of the method steps of an embodiment of the present invention.

[0057] Figure 2 This is a model architecture diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0058] The present invention is further explained below in detail with reference to the accompanying drawings so that those skilled in the art can more deeply understand the present invention and be able to implement it. However, the following reference examples are only used to explain the present invention and are not intended to limit the present invention.

[0059] like Figure 2 As shown in the figure, a graph structure is constructed based on the user's historical interaction behavior, and nodes such as users, interaction behaviors, and products are connected to form a complete graph representation. Then, the model extracts product features in the source domain, global domain, and target domain through multiple GCN layers to generate corresponding embedded representations. These embedded representations are optimized by the contrastive learning module to further improve the representation quality and distinguishing ability of the features. Then, the multi-domain features optimized by contrastive learning are integrated through feature fusion operations, and finally the recommendation results are output through the prediction module. Such an architectural design can make full use of cross-domain information, and improve the accuracy and generalization ability of the recommendation system through the GCN graph structure feature extraction and contrastive learning optimization mechanism.

[0060] like Figure 1 As shown, a cross-domain product recommendation method based on multi-layer convolutional blocks and contrastive learning includes the following steps:

[0061] S1: Data preprocessing,In order to test the effectiveness of the cross-domain recommendation method,,the data from three different fields, "Books", "CDs_and_Vinyl" and "Movies_and_TV",in Amazon's user behavior data, are used as the original data sets,and the data are preprocessed.

[0062] S2: Model training, firstly, construct the source domain sequence graph, the target domain sequence graph and the cross-domain global graph through user behavior data. Secondly, the three sequence graphs are respectively passed through a multi-layer encoder to obtain the corresponding item embedding, and finally the final intra-domain item representation embedding and inter-domain user preference embedding are obtained through a non-linear layer and a contrastive learning module.

[0063] S3: Non-linear layer and contrastive loss: First, the obtained embedding is non-linearly projected, and the SoftMax function is used to make the project embedding more informative. In the contrastive learning module, the InfoNCE function is used to calculate the loss, maximizing the mutual information between positive sample pairs and minimizing the consistency between negative sample pairs.

[0064] S4: Prediction results: After obtaining the final intra-domain item representation view and inter-domain user preference view, the final prediction results are obtained through contrastive learning modules and linking operations.

[0065] S5: Model evaluation: The model performance is evaluated by calculating the recall rate (Recall) and the normalized discounted cumulative gain (NDCG) and compared with other models.

[0066] S2 is specifically: multi-layer convolution blocks use lightweight convolution modules to stack multiple layers, and the propagation rules between each layer of convolution blocks are:

[0067]

[0068] in, is the embedding representation of each layer of users, N u is the degree of the corresponding layer User, N i is the degree of the corresponding layer Item, is a regularization term. For each layer of embedding matrix representation, there are:

[0069]

[0070] Among them, the adjacency matrix A represents the number of non-zero entries in the i-th row vector, is a symmetric normalization term.

[0071] S3 is specifically as follows: In order to make the obtained embedding representation have nonlinear characteristics, nonlinear projection is applied at the exit of each multi-layer convolution block:

[0072] h=RELU(W*Embedding(x)+b) (3)

[0073] Among them, W is the weight matrix of the activation function, Embedding(x) is the embedding vector, b is the bias, and RELU is the activation function, which is used to improve the representation ability of the model.

[0074] For the contrastive learning module, the InfoNCE function is used as the loss function:

[0075]

[0076] Among them, x i is the embedding of the sample, x j is the positive sample embedding corresponding to the sample, x k is the negative sample embedding corresponding to the sample, sim(x i ,x j ) is the cosine similarity between samples, and τ is the smoothing coefficient.

[0077] Therefore, for the embedded preference learning, we have:

[0078]

[0079] S4 is as follows: Due to the differences in the characteristics of projects in different fields, the data in the target domain is usually sparse. In order to make full use of user information in different fields, E S and EL For aggregation, there are:

[0080] E=[E S ,E L ] (6)

[0081] For the cross-domain recommendation model, the binary cross entropy function is used as the loss function:

[0082]

[0083] Among them, j is the negative sampling term, σ(.) is the sigmoid function, δ(x t+1 ) is a filler item.

[0084] Calculate the recall rate to measure the model's ability to retrieve positive (or related) samples. It represents the proportion of all truly relevant samples that the model successfully retrieves:

[0085]

[0086] TP is the number of samples correctly predicted by the model as positive, and FN is the number of samples that are actually positive but failed to be identified by the model.

[0087] Calculate the Normalized Discounted Cumulative Gain (NDCG) to measure the quality of the ranking results of the cross-domain recommendation model:

[0088]

[0089] rel is the relevance score of the i-th sample, is the relevance score of the result at the ith position in the ideal sorting case, and k is the first k results considered when calculating NDCG.

[0090] The present invention proposes a cross-domain product recommendation method based on multi-layer convolutional blocks and contrastive learning, which integrates multi-layer convolutional blocks, contrastive learning, and linking technologies, alleviates the inherent problems of data sparsity and feature differences in the recommendation system in the field of cross-domain recommendation, enhances the recommendation system's understanding of the potential similarities between products, and improves the generalization ability of the model and the accuracy of cross-domain recommendation.

[0091] The specific implementation scheme described above further describes in detail the purpose, technical scheme and beneficial effects of the present invention. It should be understood that the above is only a specific implementation scheme of the present invention and is not intended to limit the scope of the present invention. Any equivalent changes and modifications made by any technician in the field without departing from the concept and principle of the present invention should fall within the scope of protection of the present invention.

Claims

1. A cross-domain product recommendation method based on multi-layer convolutional blocks and contrastive learning, characterized in that: The following steps are involved: S1. Data preprocessing: We use the data from three different fields, "Books", "CDs_and_Vinyl" and "Movies_and_TV" in Amazon user behavior data as the original data set and preprocess the data. S2, model training: First, construct the source domain sequence graph, the target domain sequence graph and the cross-domain global graph based on the user behavior data. Second, pass the three sequence graphs through multi-layer encoders to obtain the corresponding item embeddings. Finally, the nonlinear layer and contrastive learning module are used to obtain the final intra-domain item representation embedding and inter-domain user preference embedding. S3, non-linear layer and contrast loss: First, the obtained embedding is non-linearly projected, and the SoftMax function is used to make the item embedding obtain richer information; In the contrastive learning module, the InfoNCE function is used to calculate the loss, maximizing the mutual information between positive sample pairs and minimizing the consistency between negative sample pairs; S4, prediction results: After obtaining the final intra-domain item representation view and inter-domain user preference view, the final prediction results are obtained through the contrast learning module and the link operation; S5. Model evaluation: The model performance is evaluated by calculating the recall rate Recall and the normalized discounted cumulative gain NDCG.

2. The cross-domain product recommendation method based on multi-layer convolutional blocks and contrastive learning according to claim 1, characterized in that: The step S1 comprises the following steps: S101, read the user's behavior data from the data sets in three fields, namely user ID, item ID, rating and timestamp, and delete repeated User-Item pairs; S102. To alleviate the cold start problem, discard the records with less than 5 user or item interactions, and merge the three types of user data to reduce data sparsity and improve the generalization ability of the model. S103. During the sampling process, a popularity calculation method is used to randomly sample negative samples for each user, and finally the training set, validation set, and test set are divided according to a ratio of 7:2:

1.

3. The cross-domain product recommendation method based on multi-layer convolutional blocks and contrastive learning according to claim 2 is characterized in that: The step S2 comprises the following steps: S201, encoding the user's historical behavior data to obtain an input sequence, and obtaining a cross-domain global graph, a source domain sequence graph, and a target domain sequence graph from the input sequence; S202: Input the three sequence graphs into the corresponding multi-layer encoder, and after multiple layers of convolution, obtain the source domain sequence embedding E S , target domain sequence embedding E T and global sequence embedding E; S203. In order to alleviate the problem of data sparsity in cross-domain recommendation, the obtained embeddings are subjected to comparative learning respectively, so that user preferences in different fields complement each other and learn self-supervision signals to maximize the mutual information between them.

4. The cross-domain product recommendation method based on multi-layer convolutional blocks and contrastive learning according to claim 3 is characterized in that: The step S3 comprises the following steps: S301. In the prediction module, in order to obtain nonlinear information, nonlinear projection is applied to the output of the multi-layer encoder layer of each module, and the ReLU activation function is used to make the embedding have nonlinear characteristics, thereby improving the representation ability of the model; S302. In the contrastive learning module, InfoNCE is used as the loss function to maximize the similarity between the positive sample and the target and minimize the similarity between the negative sample and the target, so as to help the model learn effective feature representation and back-propagate the learned parameters.

5. The cross-domain product recommendation method based on multi-layer convolutional blocks and contrastive learning according to claim 4 is characterized in that: The step S4 comprises the following steps: S401, finally get the source domain sequence embedded in E S and the target domain sequence embedding E T Afterwards, due to the inconsistency of project features in different fields, the data in the target domain will be relatively sparse, and a link operation is used to aggregate the embedding representations of the source and target domains to fully utilize user information in different fields; S402. In the final prediction module, the binary cross entropy function is used as the loss function of the model. The loss is calculated by measuring the difference between the predicted probability distribution and the true label, and the model parameters are updated to enable the model to output more accurate prediction results.

6. The cross-domain product recommendation method based on multi-layer convolutional blocks and contrastive learning according to claim 2, characterized in that: Multi-layer convolution blocks use lightweight convolution modules to stack multiple layers. The propagation rules between each layer of convolution blocks are: in, is the embedding representation of each layer of users, N u is the degree of the corresponding layer User, N i is the degree of the corresponding layer Item, is a regularization term. For each layer of embedding matrix representation, there are: Among them, the adjacency matrix A represents the number of non-zero entries in the i-th row vector, is a symmetric normalization term.

7. The cross-domain product recommendation method based on multi-layer convolutional blocks and contrastive learning according to claim 3, characterized in that: In order to make the obtained embedding representation have nonlinear characteristics, nonlinear projection is applied at the exit of each multi-layer convolution block: h=RELU(W*Embedding(x)+b) (3) Among them, W is the weight matrix of the activation function, Embedding(x) is the embedding vector, b is the bias, and RELU is the activation function to improve the representation ability of the model; For the contrastive learning module, the InfoNCE function is used as the loss function: Among them, x i is the embedding of the sample, x j is the positive sample embedding corresponding to the sample, x k is the negative sample embedding corresponding to the sample, sim(x i ,x j ) is the cosine similarity between samples, τ is the smoothing coefficient; Therefore, for the embedded preference learning, we have:

8. The cross-domain product recommendation method based on multi-layer convolutional blocks and contrastive learning according to claim 4, characterized in that: Since the characteristics of projects in different fields are different, the data in the target domain is usually sparse. In order to make full use of user information in different fields, the E S and E L For aggregation, there are: And=[And S ,AND L ] (6) For the cross-domain recommendation model, the binary cross entropy function is used as the loss function: Among them, j is the negative sampling term, σ(.) is the sigmoid function, δ(x t+1 ) is a filler item.

9. The cross-domain product recommendation method based on multi-layer convolutional blocks and contrastive learning according to claim 5, characterized in that: Calculating the recall rate Recall represents the proportion of all truly relevant samples that the model successfully retrieves: Among them, TP is the number of samples correctly predicted by the model as positive, and FN is the number of samples that are actually positive but failed to be identified by the model; Calculate the normalized discounted cumulative gain NDCG to measure the quality of the ranking results of the cross-domain recommendation model: Among them, rel is the relevance score of the i-th sample, is the relevance score of the result at the ith position in the ideal sorting case, and k is the first k results considered when calculating NDCG.

Images