Heterogeneous collaborative filtering method for multi-behavior recommendation

Abstract

The application relates to a heterogeneous collaborative filtering method for multi-behavior recommendation, belongs to the multi-behavior recommendation field, and is based on a heterogeneous collaborative filtering model HCFMR for multi-behavior recommendation and comprises the following steps: S1, obtaining user and item interaction features of a specific behavior by using an improved lightweight graph convolutional network; S2, obtaining multi-behavior dependent semantics of a user bottom layer by using a multi-head attention network; S3, setting a weight of a specific behavior for each user, and the weight is used for distinguishing the contribution degrees of different behaviors; and S4, obtaining high-order embedding representation of a node by aggregating various convolutional layers, obtaining a user preference, and then performing multi-behavior recommendation. The application reduces model complexity, enriches user and item embedding representation, improves recommendation quality, and improves the explainability of the model.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of multi-behavior recommendation, and relates to a heterogeneous collaborative filtering method for multi-behavior recommendation. BACKGROUND

[0002] To deal with the predicament of information overload, a recommendation system emerges as the times require and has become an important tool for assisting users in decision-making. Traditional recommendation systems often focus on one type of interaction data (such as purchase), but in actual recommendation, multiple types of interaction data (such as click browsing, adding to the shopping cart, collecting, purchasing, etc.) can be used to build a fine-grained recommendation system and alleviate the data sparsity problem.

[0003] A multi-behavior recommendation system is based on multiple interaction data of a user. In early research, matrix decomposition and neural network methods are used to improve the recommendation effect, but they only use first-order interaction information and ignore the large amount of high-order heterogeneous interaction contained in multi-behavior data. Recently, some methods based on graph convolution network (GCN) model the user and item interaction on a bipartite graph, making up for the deficiency that early methods cannot capture high-order interaction information, but these methods follow the typical collaborative filtering idea and only encode the high-order heterogeneous paths of users and items, ignoring some fine-grained interaction semantics at the bottom of users and items.

[0004] In actual e-commerce recommendation scenarios, there are multiple behavior data such as click browsing, collecting, adding to the shopping cart, and purchasing. Each type of behavior can reflect user preferences from different angles. Some users may add the goods they are interested in to the shopping cart for comparison before purchasing, and other users may purchase after click browsing. The reason for the behavior dependence may be closely related to user characteristics (consumption ability, shopping habits) and item attributes (price). In the process of modeling user preferences, it is beneficial to integrate the dependence of users on different behaviors into the representation of users and items to improve the recommendation effect. In addition, among the multiple behaviors of users, purchase is obviously a stronger signal than click browsing. If every behavior is treated uniformly, it is obviously inappropriate. Therefore, it is necessary to represent the strength of different behaviors from the data when modeling the multi-behavior recommendation model. Some previous work always ignores these considerations, which greatly reduces the performance of the recommendation system. SUMMARY

[0005] In view of this, the purpose of the present application is to provide a multi-behavior recommendation model HCFMR (Heterogeneous collaborative filtering for multi-behavior recommendation).

[0006] To achieve the above object, the present application provides the following technical solutions:

[0007] A heterogeneous collaborative filtering method for multi-behavior recommendation based on a heterogeneous collaborative filtering model HCFMR for multi-behavior recommendation, comprising the following steps:

[0008] S1: obtaining user and item interaction features of a specific behavior by using an improved lightweight graph convolutional network;

[0009] S2: obtaining user underlying multi-behavior dependent semantics by using a multi-head attention network;

[0010] S3: setting a weight of a specific behavior for each user for distinguishing the contribution degree of different behaviors;

[0011] S4: obtaining high-order embedding representation of nodes by aggregating each convolutional layer to obtain user preferences and thus to perform multi-behavior recommendation.

[0012] Further, the step S1 of obtaining user underlying multi-behavior dependent semantics by using a multi-head attention network specifically comprises:

[0013] Based on a lightweight graph convolutional network Light GCN, the node and the relationship are jointly embedded to obtain respective embedding representations, and the representations of the node and the relationship are updated in a mutually enhanced manner, and the specific representations of the user and the relationship are as formulas (1) and (2):

[0014]

[0015]

[0016] In the formula, (i, r k )∈N(u) represents a set of items i interacting with the user u under the kth relationship, is a normalization operation; is a layer-specific trainable parameter matrix; is a combination function for integrating the relationship embedding representation into the node representation to enhance the node representation;

[0017] The neighbor features under a specific relationship are respectively aggregated, and the neighbor features of different orders of the user are obtained through L-order propagation of the graph convolutional network The relationship embedding representation contains the interaction features of the specific behavior and the underlying multi-behavior dependent semantics

[0018] Further, the step S2 of obtaining user underlying multi-behavior dependent semantics by using a multi-head attention network specifically comprises:

[0019] S21: define the query, key, value conversion matrix Q, K, for the relationship k and the projection between the embedded representations;

[0020] S22: scale the query vector and all the key vectors to get the correlation weight between the relationship embeddings and normalize it to get

[0021]

[0022] S23: weight the value vector based on the learned correlation weight of the specific head to obtain the multi-behavior dependency semantics of the user and the item

[0023]

[0024] where H represents the number of heads of the multi-head attention network, and || represents the concatenation operation;

[0025] S24: add the interaction features of the specific behavior to the underlying multi-behavior dependency semantics to obtain the final relationship embedding representation

[0026]

[0027] Further, step S3 sets a specific behavior weight β for each user uk As follows:

[0028]

[0029] where W k is the weight of the k-th behavior of user u, n uk is the number of items interacted by user u in the k-th behavior, ∑ m∈N(r) n um is the total number of items interacted by user u and ∑ m∈N(r) β uk = 1.

[0030] Further, step S4 obtains the high-order embedding representation of the node by aggregating each convolution layer to obtain the preference of the user and thus performs multi-behavior recommendation, specifically including:

[0031] Initial embedding First, it is converted into a sparse vector through one-hot encoding, and then embedded into a smaller dimensional space through multi-layer propagation, so that the user, item and relationship obtain different levels of embedding representation;

[0032] Obtain the user embedding representation for each behavior Then, based on the behavioral weight β uk right Perform a weighted summation to obtain the user representation of layer L. The L-th layer represents the sum of the embeddings of each behavior:

[0033]

[0034]

[0035] The final representation is obtained by combining the embedding representations of different layers:

[0036]

[0037] Each layer in the formula has a uniform hyperparameter weight.

[0038] Estimate the probability y(k) that user u will interact with item i under the Kth action. u,i for:

[0039]

[0040] Furthermore, the non-sampling loss function for the Kth behavior is:

[0041]

[0042] In the formula Let B represent the interaction items of user u under the Kth behavior, B represent a batch of users, and I represent the set of items. The complexity of formula (11) is O((|B|+|I|)d 2 +|I k+ |d);

[0043] HCFMR employs a multi-task training approach, and its objective function is:

[0044]

[0045] In the formula λ k These are hyperparameters specified for different types of data, used to control the impact of the Kth behavior on multi-task training; L2 regularization is used to prevent overfitting, and μ is the regularization coefficient.

[0046] Furthermore, mini-batch Adam is used as an optimizer to optimize the objective function.

[0047] Furthermore, message dropping and node dropping methods are employed to prevent overfitting of the neural network.

[0048] The beneficial effects of the present application are:

[0049] (1) By improving the lightweight graph convolutional network for multi-behavior recommendation, the model complexity is reduced compared with the previous graph convolution-based multi-behavior recommendation.

[0050] (2) By introducing a multi-head attention mechanism to grasp the dependency between different behaviors of users and items, the embedding of users and items is enriched, and the recommendation quality is improved.

[0051] (3) In the fusion of user representation, the importance of behavior and the number of user-item interactions are used to learn the weight of the user under different behaviors, improving the interpretability of the model.

[0052] Other advantages, objects and features of the present application will be in part apparent and in part pointed out hereinafter in the specification, as it will be understood to those skilled in the art upon reading the following specification and by virtue of the appended claims, or can be learned by practice of the application. The objects and other advantages of the present application can be realized and attained by the methods and instrumentalities particularly pointed out in the following description. BRIEF DESCRIPTION OF DRAWINGS

[0053] In order to make the purposes, technical solutions and advantages of the present application clearer, the preferred detailed description of the present application will be made below in combination with the drawings, in which:

[0054] Figure 1 The framework diagram of the multi-behavior recommendation-based heterogeneous collaborative filtering model of the present application;

[0055] Figure 3 The influence of the number of convolutional layers on the model;

[0056] Figure 2 The multi-head attention network structure diagram;

[0057] Figure 4 The performance diagram of different graph convolutions;

[0058] Figure 5 The influence diagram of auxiliary behavior data;

[0059] Figure 6 The influence diagram of user behavior weight and attention;

[0060] Figures 3-6 (a) and (b) in the above table are the corresponding experimental diagrams of the experimental data sets Beibei and Taobao, respectively. DETAILED DESCRIPTION

[0061] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0062] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0063] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0064] Please see Figure 1 This invention proposes a heterogeneous collaborative filtering model HCFMR (Heterogeneous Collaborative Filtering for Multi-Behavior Recommendation) based on multi-behavior recommendation, which represents the embedding of users (items, relationships) as follows: The user-project interaction matrix is ​​Y = {Y} (1) Y (2) ...Y (k) Let K represent the total number of interaction types. If the user interacts with the project y(k) under the Kth behavior, then... u,i =1, otherwise 0, where y(k) u,i ∈Y (k)In multi-behavior recommendation, a target behavior most relevant to user preferences is often set (such as purchasing in e-commerce), while other behaviors are set as auxiliary behaviors (such as clicking to browse, adding to cart, etc.). The task of HCFMR is to estimate the probability of user interaction with items under the target behavior.

[0065] HCFMR consists of three modules: 1) Embedding propagation module, which obtains the embedding representation vectors of users, items, and relationships at each level. 2) Multi-behavior prediction module, which estimates the probability that user u will interact with item i under the Kth behavior. 3) Multi-behavior optimization module, which optimizes model parameters through advanced non-sampling learning methods.

[0066] HCFMR is a multi-relationship graph The model above, in which These represent the sets of nodes, edges, and relationships, respectively. This model leverages the advantages of Light GCN neighborhood aggregation and makes some improvements. It obtains their respective embedding representations by jointly embedding nodes and relationships. The representations of nodes and relationships are updated in a mutually reinforcing manner. The update method for users is similar to that for items. The specific representations of users and relationships are shown in formulas (1) and (2):

[0067]

[0068]

[0069] In the formula (i,r) k )∈N(u) represents the set of items i that interact with user u under the k-th relation. It is a normalization operation that can prevent the embedding size from increasing with the number of graph convolutional layers. It is a layer-specific trainable parameter matrix that projects all relations into the same embedding space as the nodes and allows them to be used in the next GCN layer. It's a composition function that integrates relation embeddings into node representations, enhancing them. Partial derivatives can be computed using the automatic differentiation package in PyTorch or TensorFlow. Since users have diverse behavioral data, neighbor features under specific relations are aggregated separately, and L-order propagation through a graph convolutional network yields neighbor features of different orders for the user. Relational embeddings represent interaction features that include specific behaviors. With underlying multi-behavior dependency semantics

[0070] like Figure 2As shown, since multi-head attention has a strong ability to mine data correlations, it is introduced into multi-behavior recommendation systems to capture the potential dependency semantics between user and item behaviors. First, the query, key, and value transformation matrices Q, K, and ... are defined. Used for relation k and Projection between embedded representations, query vector With key vector The relevance weights between relation embeddings are obtained by scaling the query vector and all key vectors into a dot product. Finally, it is normalized to obtain... The specific representation is as follows:

[0071]

[0072] Next, the value vectors are weighted and concatenated based on the relevance weights of specific heads learned from learning to obtain multi-behavioral dependency semantics for users and items.

[0073]

[0074] In the formula, H represents the number of heads in the multi-head attention network, || represents the concatenation operation, and finally, the interaction features of specific behaviors are added to the underlying multi-behavior dependency semantics to obtain the final relation embedding representation.

[0075]

[0076] In multi-behavior recommendation, it's necessary to consider the contribution of different behaviors to user preference prediction. For example, in e-commerce recommendations, purchasing behavior obviously contributes the most, but this isn't always the case. When a user has few purchase records, other behaviors may play a greater role. Therefore, the model needs to automatically learn the contribution of different behaviors based on different users. HCFMR assigns different behavior weights β to each user. uk The expression is as follows:

[0077]

[0078] In the formula W k It is the weight of user u's Kth behavior, n uk ∑ is the number of items that user u interacts with under the k-th behavior. m∈N(r) n um It is the total number of items that user u interacts with and ∑ m∈N(r) β uk =1.

[0079] The user embedding representation of each behavior is obtained through the embedding propagation layer. Then, based on the behavioral weight β uk right Perform a weighted summation to obtain the user representation of layer L. The L-th layer represents the sum of the embeddings of each behavior:

[0080]

[0081]

[0082] Multi-behavior prediction module: HCF-MR initial embedding First, the data is converted into sparse vectors using one-hot encoding, and then embedded into a low-dimensional space. After multiple propagation layers, users, items, and relationships obtain embedding representations at different levels. These different levels of embedding representations represent information converged from different orders. Combining the embedding representations from different levels yields the final representation.

[0083] Each layer in the formula is assigned a uniform weight. This is a hyperparameter that enables the model to achieve better performance. It estimates the probability y(k) that user u will interact with item i under the Kth action. u,i for:

[0084]

[0085] Multi-behavior optimization module: Negative sampling has become the preferred optimization method in most works due to its high training efficiency and ease of implementation. However, its instability and high randomness cannot be ignored. If negative sampling is performed on multi-feedback data, it samples a negative example for each interaction. The number of negative examples generated in multi-feedback data is approximately K times that of single-feedback data (K is the number of behavior types). This introduces a significant increase in randomness, making model optimization difficult. Non-sampling methods consider all data, avoiding these problems and ensuring better model robustness. The non-sampling loss function for the Kth behavior is:

[0086]

[0087] In the formula Let B represent the interaction items of user u under the Kth behavior, B represent a batch of users, and I represent the set of items. The complexity of formula (11) is O((|B|+|I|)d 2 +|I k+ |d) In order to learn parameters from users' multi-behavioral data, HCFMR adopts a multi-task training approach, and its objective function is:

[0088]

[0089] In the formula λ kThese are hyperparameters specified for different types of data, designed to control the impact of the Kth behavior on multi-task training. L2 regularization is used to prevent overfitting, with μ being the regularization coefficient. To optimize the objective function, mini-batch Adam is chosen as the optimizer, its main advantage being that the learning rate can adapt during the training phase. Dropout is an effective solution to prevent overfitting in neural networks; HCFMR employs two dropout methods: message dropout and node dropout.

[0090] Experiments were conducted using the Beibei and Taobao datasets. Both datasets contained data on three types of actions: browsing, adding to cart, and purchasing. Purchasing was the target recommended behavior. Specific data is shown in Table 1. Users and items with fewer than 5 purchases were filtered out from both datasets. A user's last purchase record was used as test data, the second-to-last record as validation data, and the remaining data was used for training.

[0091] Table 1

[0092]

[0093] The model employs two commonly used evaluation metrics: Hit Ratio (HR) and Normalized Discounted Cumulative Gain (NDCG). HR emphasizes recommendation accuracy, measuring whether an item in the test set is in the user's TOP-K recommendation list. NDCG emphasizes the importance of the order of the hit item in the TOP-K recommendation list; that is, the higher the position of the hit item in the recommendation list, the better the recommendation quality.

[0094] To evaluate the performance of HCFMR, the baselines for comparison were divided into two categories: single-behavior recommendation models and multi-behavior recommendation models. The single-behavior recommendation models compared included:

[0095] (1) NCF: A neural collaborative filtering model with multiple implementation methods;

[0096] (2) Light GCN: A simplified graph collaborative filtering model of GNN with high performance in single-behavior recommendation;

[0097] Multi-behavior recommendation models include:

[0098] (1) NMTR: By applying multiple NCF modules to a multi-task learning framework, and by cascading multiple behaviors to capture more subtle information;

[0099] (2)SGCNMB: Multi-behavior recommendation is achieved by applying multiple Light GCN modules to a multi-task learning framework;

[0100] (3) EHCF: It associates the prediction of each behavior through transfer learning and uses a non-sampling method for multi-behavior recommendation;

[0101] (4)GHCF [7] By improving the traditional GCN, the model combines user item embedding and relation embedding to perform multi-behavior prediction, and incorporates advanced non-sampling learning to achieve high performance.

[0102] In the experiment, the HCFMR model used a grid search method to search for optimal parameters on the validation set and evaluated them on the test set. All baselines were initialized according to the parameters in the corresponding papers to achieve optimality. The epoch was set to 500, the batch-size to 256, the embedding dimension d to 64, and the learning rate to 0.001. The number of graph convolutional layers on the two datasets was set from [1, 2, 3, 4, 5]. In the multi-head attention network, the number of heads H was set to 2, and the length of the recommendation list was [10, 50, 100]. The parameters λ1, λ2, and λ3 in the multi-task loss function were adjusted from [0, 1 / 6, 2 / 6, 3 / 6, 4 / 6, 5 / 6, 1], and λ1 + λ2 + λ3 = 1. Determining two of them allows us to obtain the third. The optimal combination in HCF-MR is [1 / 6, 4 / 6, 1 / 6].

[0103] Table 2 shows the results of the comparative experiments on the two datasets, and the detailed analysis is as follows:

[0104] Table 2

[0105]

[0106] Table 2 shows the results of each model on the two datasets, and it can be seen that:

[0107] (1) HCF-MR outperforms the best baseline GHCF on all metrics across both datasets, achieving average improvements of 4.07% and 6.4% on the Beibei and Taobao datasets, respectively. This is mainly attributed to HCFMR's modeling of user dependence and contribution to different behaviors. SGCNMB performs bipartite graph modeling for each type of data and uses Light GCN for neighborhood feature aggregation. Compared to HCFMR, it lacks explicit modeling of higher-order heterogeneous paths for users and items, and also fails to focus on behavioral dependence and importance, greatly limiting the role of multi-behavioral data. This results in insufficient ability to extract fine-grained semantics of users and items, leading to poor performance in user preference prediction. The neural network model NMTR and the matrix factorization model EHCF, which only utilize first-order interaction information of users and items, are clearly inferior to HCFMR and GHCF in recommendation performance. This indicates that graph convolution-based methods can fully utilize the higher-order neighbors of users and items, and deeply mining higher-order collaborative information is more advantageous for predicting user preferences. The above demonstrates the effectiveness of HCFMR in introducing multi-head attention to explore user dependence on different behaviors and setting different behavior weights to distinguish the contribution of different behaviors.

[0108] (2) All multi-behavior recommendation models outperformed single-behavior recommendation models, demonstrating the benefits of introducing multi-behavior data for improving recommendation performance. This is because different behaviors are complementary and reflect the interaction patterns between users and items from different perspectives. For users with limited target behavior data, auxiliary behavior data can ensure they achieve better recommendation results. The multi-behavior recommendation models NMTR and SGCNMB are improvements on the single-behavior recommendation models NCF and LightGCN, respectively, further illustrating the advantages of introducing multi-behavior data in improving recommendation performance.

[0109] (3) For all evaluation criteria, all models outperformed the Taobao dataset on the Beibei dataset. This is attributed to the fact that the latter has greater sparsity, which limits the model's ability to extract information.

[0110] To evaluate the model complexity of HCFMR, the average training time of multiple multi-behavior recommendation models on the two datasets was recorded under the same experimental conditions. Table 3 shows that the average training time of HCFMR is 9.1% and 16.5% lower than the best baseline GHCF on the Beibei and Taobao datasets, respectively. This is mainly because HCFMR uses a lightweight graph convolutional network for neighborhood aggregation iteration, significantly reducing the number of model parameters and lowering the training difficulty. This also demonstrates the effectiveness of improving the lightweight graph convolutional network. The three models take longer on the Taobao dataset than on the Beibei dataset, primarily because the Taobao dataset has a larger number of users and items.

[0111] Table 3

[0112]

[0113] To investigate the impact of the number of graph convolution layers on the model, HCFMR was tested with different depths. The performance results for HR@100 on two datasets are as follows: Figure 3 As shown in (a) and (b) in the figure.

[0114] from Figure 3 The results show that the performance of HCFMR increases with the number of convolutional layers. On the Beibei dataset, it reaches its optimal performance with 3 convolutional layers, improving performance by 2.0% compared to a single convolutional layer. On the Taobao dataset, it reaches its optimal performance with 4 convolutional layers, improving performance by 4.6% compared to a single convolutional layer. Afterward, the model performance begins to decline with further increases in the number of convolutional layers. This indicates that stacking appropriate convolutional layers helps explore higher-order neighbor information of user items, improving recommendation performance, especially for sparse datasets. However, stacking too many convolutional layers introduces noise and weakens the model's performance.

[0115] The improvement of HCFMR is based on a lightweight graph convolutional network, while also drawing inspiration from R-GCN, a multi-relation graph neural network used to handle entity and relation embeddings in knowledge graphs.

[16] and Comp-GCN

[17] By incorporating relationships into the node embedding representation, this approach is better suited for multi-behavior recommendation. Therefore, to verify the effectiveness of the HCFMR improvement, the graph convolutional network of this invention was replaced with the two methods described above. Experimental results on two datasets are as follows: Figure 4 (a)(b) in the text.

[0116] from Figure 4 The results show that HCFMR achieves the best improvement, with improvements of 27.1% and 33.8% compared to R-GCN on the two datasets. This is because the R-GCN model assigns a parameter matrix to each relation, and as the number of relations increases, the introduction of too many relation matrices makes the model untrainable. Compared to Comp-GCN, the improvements on the two datasets are 4.0% and 5.0%. This is because HCFMR adopts an iterative update method that mutually reinforces nodes and relations, integrating relations into node representations while also considering relation representation. In multi-head attention networks, obtaining the underlying multi-behavioral dependency semantics of users also requires stronger relation representations.

[0117] To explore the impact of auxiliary behavioral data on the experiment, several variant experiments were conducted on two datasets: HCFMR-P, which only included purchase behavior data; HCFMR-PV, which included purchase and click / browse data; and HCFMR-PC, which included purchase and add-to-cart data. The results on both datasets are as follows: Figure 5 As shown in (a) and (b) in the figure.

[0118] from Figure 5 Overall, the variant model using auxiliary behavioral data significantly outperformed the model HCFMR-P, which only used target behavioral data. The HCFMR model, using all three datasets, improved performance by 22.3% and 31.8% on the Beibei and Taobao datasets, respectively, compared to HCFMR-P which only used purchase data. This indicates that introducing auxiliary behavioral data has a positive effect on the model. Adding shopping cart behavior to both datasets yielded lower results than adding click-and-browse behavior, which may be due to differences in the size of the auxiliary behavioral data.

[0119] To evaluate the effectiveness of HCFMR's user behavior weights and attention module, three variant models were proposed: HCFMR-1 (retaining only the graph convolutional network module), HCFMR-2 (using uniform behavior weights for different user behaviors), and HCFMR-3 (removing the attention module). Experimental results on two datasets are as follows: Figure 6 (a)(b) in the text.

[0120] from Figure 6 The results show that the HCFMR model without ablation outperforms the three variant models on both datasets. Compared to HCFMR-1, which only retains the graph convolutional module, the unablated HCFMR model achieves an average improvement of 2.8% and 3.3% on the Beibei and Taobao datasets, respectively, demonstrating the effectiveness of introducing user behavior weights and attention. It is noteworthy that HCFMR-2 and HCFMR-3 outperform HCFMR-1 on both datasets, with HCFMR-3 showing slightly better performance, indicating that introducing user behavior weights is more helpful in improving model performance.

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A heterogeneous collaborative filtering method for multi-behavior recommendation, characterized in that: The heterogeneous collaborative filtering model HCFMR based on multi-behavior recommendation includes the following steps: S1: Utilize an improved lightweight graph convolutional network to obtain user and item interaction features for specific behaviors, specifically including: Based on the Lightweight Graph Convolutional Network (Light GCN), nodes and relations are jointly embedded to obtain their respective embedding representations. The representations of nodes and relations are updated in a mutually reinforcing manner. The specific representations of users and relations are shown in formulas (1) and (2): (1) (2) In the formula Representing the k Under this relationship with users u Interactive projects i gather, It is a normalization operation; It is a layer-specific trainable parameter matrix; It is a composition function used to incorporate relational embeddings into node representations, thereby enhancing node representations; Neighbor features under specific relationships are aggregated separately, and then neighbor features of different orders are obtained through L-order propagation of a graph convolutional network. Relational embeddings represent interaction features that include specific behaviors. With underlying multi-behavior dependency semantics S2: Obtain the underlying multi-behavioral dependency semantics of users through a multi-head attention network, specifically including: S21: Define the query, key, and value transformation matrix , , Used for relationships and Projection between embedded representations; S22: Query vector With all key vectors Perform a scaled dot product to obtain the relevance weights between relation embeddings. And it is normalized to obtain ; (3) S23: Weight the value vectors based on the relevance weights of the learned specific heads to obtain the multi-behavioral dependency semantics of users and items. (4) In the formula, H represents the number of heads in the multi-head attention network, and || represents the splicing operation; S24: Add the interaction features of specific behaviors to the underlying multi-behavior dependency semantics to obtain the final relation embedding representation. : (5) S3: Assign a weight to a specific behavior for each user to differentiate the contribution of different behaviors; S4: By aggregating the high-order embedding representations of nodes from various convolutional layers, user preferences are obtained, thereby enabling multi-behavioral recommendations.

2. The heterogeneous collaborative filtering method for multi-behavior recommendation according to claim 1, characterized in that: Step S3 involves assigning a weight to a specific behavior for each user. as follows: (6) In the formula User u The weight of the Kth behavior, User u The number of items interacted under the k-th behavior. User u The total number of interactive items and .

3. The heterogeneous collaborative filtering method for multi-behavior recommendation according to claim 1, characterized in that: Step S4 describes obtaining high-order embedding representations of nodes by aggregating various convolutional layers to determine user preferences, thereby enabling multi-behavior recommendations. Specifically, this includes: Initial embedding , , First, it is converted into a sparse vector through one-hot encoding, and then embedded into a space with a small dimension. After multiple propagations, users, items, and relationships obtain different levels of embedded representations. Obtain the user embedding representation for each behavior Then, based on the behavior weight right Perform a weighted summation to obtain the user representation of layer L. The summation of the embedding representations of each behavior for the L-th layer item is obtained as follows: (7) (8) The final representation is obtained by combining the embedding representations of different layers: (9) Each layer in the formula has a uniform hyperparameter weight. ; Estimated users u For the project i There is a possibility of interaction under the Kth action. for: (10)。 4. The heterogeneous collaborative filtering method for multi-behavior recommendation according to claim 1, characterized in that: The non-sampling loss function for the Kth behavior is: (11) In the formula On behalf of users u For the interaction items under the Kth behavior, B represents a batch of users, I represents the set of items, and the complexity of formula (11) is ; HCFMR employs a multi-task training approach, and its objective function is: (12) In the formula These are hyperparameters specified for different types of data, used to control the impact of the Kth behavior on multi-task training; using Regularization prevents overfitting. It is the regularization coefficient.

5. The heterogeneous collaborative filtering method for multi-behavior recommendation according to claim 4, characterized in that: The objective function is optimized using mini-batch Adam as the optimizer.

6. The heterogeneous collaborative filtering method for multi-behavior recommendation according to claim 5, characterized in that: Message dropping and node dropping methods are used to prevent overfitting of neural networks.

Images