A multi-behavior e-commerce platform recommendation method based on collaborative contrastive learning

Abstract

A multi-behavior e-commerce platform recommendation method based on collaborative contrastive learning. First, a multi-behavior e-commerce user-e-commerce product interaction bipartite graph is constructed for the online e-commerce platform. Next, a multi-behavior cascaded residual network is used to capture the deep representation of the bipartite graph, and a multi-behavior attention network is used to capture the interaction representation of the bipartite graph. The representation performance of the e-commerce user-e-commerce product interaction bipartite graph is then improved through cross-layer contrastive learning, cross-behavior contrastive learning, and cross-network contrastive learning. Simultaneously, a Bayesian personalized ranking loss is calculated to improve recommendation performance. Finally, the final e-commerce user and e-commerce product representations are obtained to achieve personalized product recommendations. This method deeply mines individual behavior information and multi-behavior interaction information, and uses contrastive learning to improve the representation performance of e-commerce users and e-commerce products, which helps to generate more accurate and user-friendly recommendation results for e-commerce platforms.

Description

Technical Field

[0001] The present invention relates to the field of recommendation, and in particular to a multi-behavior e-commerce platform recommendation method based on collaborative comparative learning. Background Art

[0002] With the rapid development of e-commerce platforms, user behaviors such as clicks, adding items to carts, saving items, and purchasing have become increasingly diverse and complex. These behaviors not only reflect users' preferences for products but also reveal differences in their preferences during the decision-making process. While clicks are often exploratory and carry a degree of preference uncertainty, behaviors such as adding items to carts and purchasing more directly reflect users' product preferences and consumption intentions. However, existing recommendation methods typically focus solely on a single behavior or simply treat different behaviors as equivalent, ignoring the potential connections and inherent differences between different behaviors, which in turn affects the accuracy and personalization of recommendation results.

[0003] Furthermore, in real-world recommendation scenarios, multi-behavior data often suffers from sparsity. While purchase behavior is less abundant than click behavior, it holds greater value in identifying user preferences. Effectively utilizing this sparse yet important behavioral data and uncovering deeper connections between multiple behaviors are key challenges in the recommendation field. Furthermore, interaction data also reveals both common and individual user preferences. Mining this information can help generate more accurate recommendations that better align with user interests. Summary of the Invention

[0004] In order to solve the problem in the existing technology of ignoring the interactive information of multiple behaviors and the intrinsic difference information between different behaviors, the present invention proposes a multi-behavior e-commerce platform recommendation method based on collaborative comparative learning, which can effectively capture the deep correlation and interactive information between behaviors.

[0005] The technical specific steps adopted by the present invention to solve its technical problem are:

[0006] A multi-behavior e-commerce platform recommendation method based on collaborative contrastive learning. First, a multi-behavior e-commerce user-e-commerce product interaction bipartite graph is constructed for online e-commerce platforms. Then, a multi-behavior cascaded residual network is used to capture the deep representation of the bipartite graph, and a multi-behavior attention network is used to capture the interaction representation of the bipartite graph. Then, the representation performance of the e-commerce user-e-commerce product interaction bipartite graph is improved through cross-layer contrastive learning, cross-behavior contrastive learning, and cross-network contrastive learning. At the same time, a Bayesian personalized ranking loss is calculated to improve the recommendation performance. Finally, the final e-commerce user and e-commerce product representations are obtained to achieve personalized product recommendations.

[0007] Furthermore, the method comprises the following steps:

[0008] Step 1: Construct a multi-behavior bipartite graph G for e-commerce user-e-commerce product interactions k , where multiple behaviors include click, add to cart, purchase, etc., totaling K behaviors, k∈{1,2,…,K};

[0009] Step 2: Use the double-layer graph neural network and the residual representation of the previous behavior to obtain the node representation of the next behavior, and finally obtain the multi-behavior cascade residual network representation e p ;

[0010] Step 3: Calculate the initial representation of the multi-behavior attention network of the e-commerce platform

[0011] Step 4: Initial characterization Attention aggregation is used to obtain multi-behavior attention network representations q ;

[0012] Step 5: Calculate the cross-layer contrastive learning loss of the kth behavior in the multi-behavior cascade residual network

[0013] Step 6. Calculate the cross-layer contrastive learning loss of the k-th behavior in the multi-behavior attention network

[0014] Step 7. Calculate the total cross-layer contrastive learning loss L layer ;

[0015] Step 8: Calculate the cross-behavior contrast learning loss L of e-commerce users’ interactions with e-commerce products behavior ;

[0016] Step 9. Calculate the contrastive learning loss L between the cascaded residual network and the multi-behavior attention network network ;

[0017] Step 10: Calculate Bayesian personalized ranking loss L bpr ;

[0018] Step 11. Calculate the final loss function L = L layer +L behavior +L network +L bpr ;

[0019] Step 12: Repeat steps 2 to 11 until convergence to obtain the final node representation matrix e f ;

[0020] Step 13: Calculate the representation similarity score to achieve personalized e-commerce product recommendations for e-commerce users.

[0021] Preferably, in step 1, user behavior data of the e-commerce platform is collected, including the e-commerce user ID, e-commerce product ID and corresponding interaction behavior type involved in K behaviors, and the e-commerce user-e-commerce product bipartite graph G of the k-th behavior k ={V k ,E k}, where k∈{1,2,…,K}, V k ={U k ,I k} represents the e-commerce user and e-commerce product nodes of the bipartite graph, U k and I k are the e-commerce users and e-commerce products of the k-th behavior, E k is the interaction relationship between e-commerce users and e-commerce products for the k-th behavior. The initial representation matrices of e-commerce users and e-commerce products are e u and e i , e u The xth row of represents the representation of the xth e-commerce user, e i The y-th row of represents the representation of the y-th e-commerce product, x∈{1,…,M},y∈{1,…,N}, M and N are the total number of e-commerce users and the number of e-commerce products for all behaviors, respectively.

[0022] More preferably, in step 2, the e-commerce user-e-commerce product bipartite graph G of the k-th behavior is calculated k Node representation

[0023]

[0024] where k∈{1,…,K}, Gnn is a two-layer graph neural network, and the initial node representation of the k=1th behavior is and are the initial e-commerce user representation and e-commerce product representation respectively, then the multi-behavior cascade residual network representation of the e-commerce platform

[0025] In step 3, the initial representation of the multi-behavior attention network of the e-commerce platform is calculated

[0026]

[0027] in is the bias representation of the randomly initialized k-th behavior, k∈{1,2,…,K}, the initial node representation and They are the initial e-commerce user representation and e-commerce product representation respectively;

[0028] In step 4, the multi-behavior attention network representation of the e-commerce platform is calculated in

[0029]

[0030] is the node fusion representation of the k-th behavior, k∈{1,2,…,K}, Represents the operation of stacking node vectors of multiple behaviors into a matrix, is the attention value between the k-th behavior and all behaviors.

[0031] More preferably, in step 5, the cross-layer contrastive learning loss of the kth behavior of the multi-behavior cascade residual network of the e-commerce platform is calculated

[0032]

[0033] where k∈{1,2,…,K};

[0034] In step 6, the cross-layer contrastive learning loss of the k-th behavior of the multi-behavior attention network of the e-commerce platform is calculated.

[0035]

[0036] where k∈{1,2,…,K};

[0037] In step 7, the total cross-layer contrastive learning loss of the e-commerce platform is calculated

[0038]

[0039] where k∈{1,2,…,K};

[0040] In step 8, the cross-behavior contrastive learning loss of e-commerce users’ interactions with e-commerce products is calculated.

[0041]

[0042] Where k1 and k2 represent two different behaviors, k∈{1,2,…,K};

[0043] In step nine, the cross-network contrastive learning loss between the e-commerce platform cascade residual network and the multi-behavior attention network is calculated.

[0044]

[0045] In step 10, calculate the Bayesian personalized ranking loss of the e-commerce platform

[0046]

[0047] In step 11, the final loss of the e-commerce platform is calculated as

[0048] L=L layer +L behavior +L network +L bpr ;

[0049] In step 12, determine whether L is less than or equal to the specified threshold. If not, return to step 2; if so, end the calculation, and the final node representation matrix is

[0050]

[0051] where e f ∈(e f,u ,e f,i ), e f,u is the representation matrix of e-commerce users, e f,i It is the representation matrix of e-commerce products;

[0052] In the step 13, for the xth e-commerce user, its representation is the e-commerce user representation matrix e f,u For the yth e-commerce product in the xth row, its representation is the e-commerce product representation matrix e f,i In the yth row of , the inner product of the two is calculated as the similarity score between the xth e-commerce user and the yth e-commerce product; for the xth e-commerce user, the similarity score between the user and all e-commerce products is calculated. f,i The similarity scores between them are calculated, and the top-K e-commerce products with the highest similarity scores are taken as the recommended e-commerce products for the e-commerce user.

[0053] The technical concept of the present invention is: in the multi-behavior e-commerce recommendation method, the potential connections between multiple behaviors are obtained through a cascaded residual network, and the behavioral differences between multiple behaviors are obtained through an attention network, and multiple comparative learning is used to capture the local and global information of the interaction between users and products, thereby improving the accuracy of multi-behavior recommendations.

[0054] The beneficial effects of the present invention are: it can process the multi-behavioral interaction information of e-commerce users and e-commerce products in the e-commerce platform, and continuously optimize the personalized recommendation results by comprehensively considering the potential connections and inherent differences between the multi-behavioral information, thereby improving the click-through rate and conversion rate of the e-commerce platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 Flowchart of the multi-behavior e-commerce platform recommendation method based on collaborative comparative learning. DETAILED DESCRIPTION

[0056] The present invention will be further described below with reference to the accompanying drawings.

[0057] Reference Figure 1, a multi-behavior e-commerce platform recommendation method based on collaborative contrastive learning, including the following steps:

[0058] Step 1: Collect user behavior data from the e-commerce platform, including the e-commerce user ID, e-commerce product ID, and corresponding interaction behavior type involved in K behaviors such as click, add to cart, and purchase, and the e-commerce user-e-commerce product bipartite graph G of the k-th behavior. k ={V k ,E k}, where k∈{1,2,…,K}, V k ={U k ,I k} represents the e-commerce user and e-commerce product nodes of the bipartite graph, U k and I k are the e-commerce users and e-commerce products of the k-th behavior, E k is the interaction relationship between e-commerce users and e-commerce products for the k-th behavior. The initial representation matrices of e-commerce users and e-commerce products are e u and e i , e u The xth row of represents the representation of the xth e-commerce user, e i The y-th row represents the representation of the y-th e-commerce product, x∈{1,…,M},y∈{1,…,N}, where M and N are the total number of e-commerce users and the number of e-commerce products for all behaviors respectively;

[0059] Step 2: Calculate the bipartite graph G of the e-commerce user and e-commerce product of the k-th behavior k Node representation

[0060]

[0061] where k∈{1,…,K}, Gnn is a two-layer graph neural network, and the initial node representation of the k=1th behavior is and are the initial e-commerce user representation and e-commerce product representation respectively, then the multi-behavior cascade residual network representation of the e-commerce platform

[0062] Step 3: Calculate the initial representation of the multi-behavior attention network of the e-commerce platform

[0063]

[0064] in is the bias representation of the randomly initialized k-th behavior, k∈{1,2,…,K}, the initial node representation and They are the initial e-commerce user representation and e-commerce product representation respectively;

[0065] Step 4: Calculate the multi-behavior attention network representation of the e-commerce platform in

[0066]

[0067] is the node fusion representation of the k-th behavior, k∈{1,2,…,K}, Represents the operation of stacking node vectors of multiple behaviors into a matrix, is the attention value between the k-th behavior and all behaviors;

[0068] Step 5: Calculate the cross-layer contrastive learning loss of the k-th behavior in the multi-behavior cascade residual network of the e-commerce platform

[0069]

[0070] where k∈{1,2,…,K};

[0071] Step 6: Calculate the cross-layer contrastive learning loss of the k-th behavior in the multi-behavior attention network of the e-commerce platform

[0072]

[0073] where k∈{1,2,…,K};

[0074] Step 7: Calculate the total cross-layer contrastive learning loss of the e-commerce platform

[0075]

[0076] where k∈{1,2,…,K};

[0077] Step 8: Calculate the cross-behavior contrastive learning loss of e-commerce users’ interactions with e-commerce products

[0078]

[0079] Where k1 and k2 represent two different behaviors, k∈{1,2,…,K};

[0080] Step 9: Calculate the cross-network contrastive learning loss between the e-commerce platform cascade residual network and the multi-behavior attention network

[0081]

[0082] Step 10: Calculate the Bayesian personalized ranking loss of the e-commerce platform

[0083]

[0084] Step 11: Calculate the final loss of the e-commerce platform

[0085] L=L layer +L behavior +L network +L bpr ;

[0086] Step 12: Determine whether L is less than or equal to the specified threshold. If not, return to step 2. If so, end the calculation. The final node representation matrix is

[0087]

[0088] where e f ∈(e f,u ,e f,i ), e f,u is the representation matrix of e-commerce users, e f,i It is the representation matrix of e-commerce products;

[0089] Step 13: For the xth e-commerce user, its representation is the e-commerce user representation matrix e f,u For the yth e-commerce product in the xth row, its representation is the e-commerce product representation matrix e f,i In the yth row of , the inner product of the two is calculated as the similarity score between the xth e-commerce user and the yth e-commerce product; for the xth e-commerce user, the similarity score between the user and all e-commerce products is calculated. f,i The similarity scores between them are calculated, and the top-K e-commerce products with the highest similarity scores are taken as the recommended e-commerce products for the e-commerce user.

[0090] In this embodiment, the e-commerce platform obtains the e-commerce user's representation information based on the e-commerce user's multi-behavior historical interaction information, mainly the e-commerce user's click, add to shopping cart, favorite, purchase and other behaviors of the e-commerce product, and finds the most similar products from the e-commerce products based on the information as recommended products for the e-commerce user.

[0091] The embodiments of this specification are merely examples of implementations of the invention and are provided for illustrative purposes only. The scope of protection of the present invention should not be considered limited to the specific embodiments described in these embodiments. The scope of protection of the present invention also extends to equivalent technical means that can be conceived by a person of ordinary skill in the art based on the invention.

Claims

1. A multi-behavior e-commerce platform recommendation method based on collaborative contrastive learning, characterized in that: First, a multi-behavior bipartite graph of e-commerce user-product interactions is constructed for an online e-commerce platform. Next, a multi-behavior cascaded residual network is used to capture the deep representation of the bipartite graph, and a multi-behavior attention network is used to capture the interaction representation of the bipartite graph. Finally, the representation performance of the bipartite graph of e-commerce user-product interactions is improved through cross-layer contrastive learning, cross-behavior contrastive learning, and cross-network contrastive learning. Furthermore, a Bayesian personalized ranking loss is calculated to improve recommendation performance. Finally, the final e-commerce user and product representations are obtained to enable personalized product recommendations. The method comprises the following steps: Step 1: Construct a multi-behavior bipartite graph of e-commerce user-e-commerce product interactions , including multiple behaviors such as click, add to cart, purchase, etc. behavior, ; Step 2: Use the two-layer graph neural network and the residual representation of the previous behavior to obtain the node representation of the next behavior, and finally obtain the multi-behavior cascade residual network representation ; Step 3: Calculate the initial representation of the multi-behavior attention network of the e-commerce platform ; Step 4: Initial characterization Using attention aggregation to obtain multi-behavior attention network representation ; Step 5: Calculate the multi-behavior cascade residual network part Cross-layer contrastive learning loss for each behavior ; Step 6: Calculate the multi-behavior attention network part Cross-layer contrastive learning loss for each behavior ; Step 7: Calculate the total cross-layer contrastive learning loss ; Step 8: Calculate the cross-behavior contrastive learning loss of e-commerce users’ interactions with e-commerce products ; Step 9. Calculate the contrastive learning loss between the cascaded residual network and the multi-behavior attention network ; Step 10: Calculate Bayesian personalized ranking loss ; Step 11. Calculate the final loss function ; Step 12: Judgment Is it less than or equal to the specified threshold? If not, return to step 2; if so, end the calculation and obtain the final node representation matrix ; Step 13: Calculate the representation similarity score to achieve personalized e-commerce product recommendations for e-commerce users; In the step 2, calculate the E-commerce user-e-commerce product bipartite graph with different behaviors Node representation in , , It is a two-layer graph neural network. The initial The node representation of a behavior is , and are the initial e-commerce user representation and e-commerce product representation respectively, then the multi-behavior cascade residual network representation of the e-commerce platform ; In step 3, the initial representation of the multi-behavior attention network of the e-commerce platform is calculated in The first randomly initialized Biased representation of behavior, , the initial node representation , and They are the initial e-commerce user representation and e-commerce product representation respectively; In step 4, the multi-behavior attention network representation of the e-commerce platform is calculated ,in For the Node fusion representation of each behavior, , Represents the operation of stacking node vectors of multiple behaviors into a matrix, For the The attention value between a behavior and all behaviors.

2. The method for recommending multiple e-commerce platforms based on collaborative comparative learning according to claim 1, characterized in that: In step 1, user behavior data of the e-commerce platform is collected, including The e-commerce user ID, e-commerce product ID and corresponding interaction behavior type involved in each behavior, E-commerce user-e-commerce product bipartite graph with different behaviors ,in , Represents the e-commerce user and e-commerce product nodes of the bipartite graph, and Respectively The set of e-commerce users and e-commerce products with different behaviors, For the The interaction relationship between e-commerce users and e-commerce products is the initial representation matrix of e-commerce users and e-commerce products. and , No. Row represents the Characterization of e-commerce users, No. Row represents the The representation of an e-commerce product, , , and They are the total number of e-commerce users and e-commerce products for all behaviors respectively.

3. The multi-behavior e-commerce platform recommendation method based on collaborative comparative learning according to claim 1 or 2, characterized in that: In the step 5, the multi-behavior cascade residual network part of the e-commerce platform is calculated. Cross-layer contrastive learning loss for each behavior in ; In step 6, the calculation of the multi-behavior attention network part of the e-commerce platform Cross-layer contrastive learning loss for each behavior in ; In step 7, the total cross-layer contrastive learning loss of the e-commerce platform is calculated in ; In step 8, the cross-behavior contrastive learning loss of e-commerce users’ interactions with e-commerce products is calculated. in and Indicates two different behaviors. ; In step nine, the cross-network contrastive learning loss between the e-commerce platform cascade residual network and the multi-behavior attention network is calculated. In step 10, calculate the Bayesian personalized ranking loss of the e-commerce platform 4. The method for recommending multiple e-commerce platforms based on collaborative comparative learning according to claim 1 or 2, characterized in that: In step 11, the final loss of the e-commerce platform is calculated as In the step 12, it is determined Is it less than or equal to the specified threshold? If not, return to step 2; if so, end the calculation, and the final node representation matrix is in , is the representation matrix of e-commerce users, It is the representation matrix of e-commerce products; In the step thirteen, for the e-commerce users, whose representation is the e-commerce user representation matrix No. Row, for e-commerce products, which are represented by the e-commerce product representation matrix No. Calculate the inner product of the two as the first e-commerce users and Similarity score of e-commerce products; e-commerce users, calculate their relationship with all e-commerce products The similarity score between them is the highest one. e-commerce products as recommended e-commerce products for the e-commerce user.

Images