CN117421661A - A group recommendation method based on counterfactual enhanced graph convolutional network - Google Patents

Abstract

The invention relates to the technical field of group recommendation, and provides a group recommendation method of a graph rolling network based on inverse fact enhancement, which comprises the following steps: constructing a group user hypergraph and a group project bipartite graph; the hypergraph of the group users is mined and the high-order information between the groups and the users is obtained to obtain user-level group representation; disturbance and convolution are carried out on the hypergraph of the group user, and a counterfactual group representation is obtained; mining high-order information between the group and the project for the group project bipartite graph and the user-level group representation, obtaining group-level group representation and project representation, and weighting and fusing the user-level group representation and the group-level group representation to obtain final group representation; training the hierarchical graph convolutional network based on the inverse fact loss, the group loss and the user loss obtained by the representation layer-by-layer training, and obtaining a recommended item list through the obtained prediction model. The invention can accurately model the characteristics of the group, the user and the project explicitly, and also enhances the interpretability of group recommendation and the adaptability of the model.

Description

A group recommendation method based on counterfactual enhanced graph convolutional network

Technical field

The present invention relates to the technical field of group recommendation, and in particular to a group recommendation method based on counterfactual enhanced graph convolution network.

Background technique

As a powerful tool to solve the problem of information overload, recommendation systems are widely used in online information systems such as e-commerce platforms, social media websites, and news portals to help users choose from various options in daily life. Recommendation systems collect users’ past preferences and generate appropriate project recommendations for users. An effective recommendation scheme can not only increase traffic and profits for service providers, but also help users discover items of interest more easily.

With the popularity of social media, online group activities have become extremely common in current social networks. However, traditional personalized recommendation algorithms are mainly designed for the preferences of individual users and are difficult to meet the needs of group users. In a group, the interests and preferences of each member may have an impact on the final group decision. Therefore, there is an urgent need for a recommendation system specifically designed for group users, that is, a group recommendation system. This kind of system takes into account the interaction and influence of members within the group and can more accurately provide personalized recommendation services to the group.

In the group recommendation process, it is a crucial task to effectively aggregate the preferences of all members in the group to form group preferences. Traditional aggregation methods are usually based on predefined heuristic rules, including fairness strategy, minimum pain strategy, maximum satisfaction strategy, etc. Although these data-independent static strategies can satisfy the formation of group preferences to a certain extent, in fact group decision-making is a complex dynamic process that needs to take into account the weight of each member. The development of neural attention mechanisms provides a flexible way to solve this problem. Recently, complex graph neural networks and innovative hypercube structures, combined with techniques such as contrastive learning, have made significant progress in group recommendation methods, successfully solving specific challenges and existing problems.

However, current group recommendation methods are usually built on a statistical framework, mainly considering the statistical relationship between groups and items, ignoring potential causal relationships. This limitation leads to the final recommendation results being less accurate, and counterfactual learning As a method in causal inference, its mining of causality has great application prospects in group recommendation. However, as the data sparsity changes, the performance of existing methods is not good. Data adaptability is also insufficient.

Contents of the invention

The present invention aims to solve at least one of the technical problems existing in the related art. To this end, the present invention provides a group recommendation method based on counterfactual enhanced graph convolution network.

The present invention provides a group recommendation method based on counterfactual enhanced graph convolution network, including:

S1: Preprocess the public group data, obtain the negative example data set, and construct the group user hypergraph and group project bipartite graph;

S2: Convolve the group user hypergraph to mine high-order information between groups and users, and obtain user-level group representation; perform perturbation on the group user hypergraph, and perform perturbation on the group user hypergraph. The graph is convolved to obtain counterfactual group representation;

S3: Convolve the group-item bipartite graph and the user-level group representation to mine high-order information between groups and items, obtain group-level group representation and group-level item representation, and pass the residual gate A control mechanism weightedly fuses the user-level group representation and the group-level group representation to obtain the final group representation;

S4: Obtain counterfactual loss based on the counterfactual group representation, user-level group representation and initial item representation, obtain group loss based on the final group representation and the group-level item representation, based on the initial user representation and The initial project indicates that the user loss is obtained, and the hierarchical graph convolution network is trained for multiple rounds according to the counterfactual loss, the group loss and the user loss to obtain a group prediction model;

S5: Input the list of items to be recommended into the prediction model, obtain a prediction score, and obtain a list of recommended items based on the ranking of the prediction score.

According to a group recommendation method based on counterfactual enhanced graph convolution network provided by the present invention, in step S1, the group user hypergraph is constructed based on group user relationship data, and the group item bipartite graph is constructed based on Group project interaction data construction.

According to a group recommendation method based on counterfactual enhanced graph convolution network provided by the present invention, in step S1, the negative example data includes an initial representation set constructed based on the ID embedding vector lookup table, and the initial representation set includes Initial group representation, initial user representation, and initial project representation.

According to a group recommendation method based on counterfactual enhanced graph convolution network provided by the present invention, in step S2, the expression of the user-level group representation is:

in, Represented as a user-level group,/> is the total number of convolutional layers of the hypergraph convolutional network,/> is the convolution layer index value of the hypergraph convolution network,/> For the first/> Aggregated information represented by nodes updated by convolutional layers.

According to a group recommendation method based on counterfactual enhanced graph convolution network provided by the present invention, in step S2, the calculation formula for perturbing the group user hypergraph is:

in, is the perturbed group user hypergraph correlation matrix,/> is the group user hypergraph correlation matrix,/> is the indicator function,/> is a trainable mask matrix,/> Threshold specified for indicator function decision entries.

According to a group recommendation method based on counterfactual enhanced graph convolution network provided by the present invention, the expressions of the group-level group representation and group-level item representation obtained in step S3 are:

in, For group-level group representation,/> Represented as a group-level item,/> For the first/> The node representation obtained by the convolutional layer, Average of node embedding representations obtained for all convolutional layers.

According to a group recommendation method based on counterfactual enhanced graph convolution network provided by the present invention, the expression of the final group representation in step S3 is:

in, is the final group representation,/> Represents the weight of user-level groups,/> is the sigmoid function,/> is the first gating coefficient,/> is the second gating coefficient,/> is the third weight coefficient.

According to a group recommendation method based on counterfactual enhanced graph convolution network provided by the present invention, in step S4, the loss used for the multi-round training of the hierarchical graph convolution network includes:

in, is the counterfactual loss,/> For group losses,/> Loss for users,/> For the group project training set,/> Data from the training set for the group project,/> Training set for user projects,/> Data from the training set for user projects,/> is the item index value,/> is the sampling negative example corresponding to the item that interacts with the group,/> is the sampling negative example corresponding to the item that interacts with the user,/> Predict the score for the fact,/> is the counterfactual prediction score,/> represents the predicted score for the final group,/> Represents predicted scores for group-level items,/> The user initially indicates the predicted score,/> Predict the score for the initial representation of the project.

The present invention provides a group recommendation method based on counterfactual enhanced graph convolution network, which aims at the shortcomings of the current group recommendation model in mining causality and adapting to different sparsity data to mine groups. The causal relationship between members and item recommendations, explore the basic principles behind group recommendations, and improve the model's adaptability to data with different sparsities.

The present invention proposes a group recommendation method CF-hGCN based on counterfactual enhanced hierarchical graph convolution network, which is mainly used in group recommendation tasks. By integrating graph counterfactual learning technology and Hypergraph convolutional network, CF-hGCN can deeply explore the causal relationship between group members and item recommendations. This method enables CF-hGCN to accurately identify key members that affect group recommendations and provides the basic principle of group recommendations. Provides in-depth explanations.

CF-hGCN adopts an innovative hierarchical network structure, which sequentially calculates the representations of groups, users and projects through the counterfactual hypergraph learning module and the bipartite graph learning module. This design enables it to operate at different data sparsities. Excellent performance in the scene. Among them, the hypergraph learning module can effectively capture the preferences of members for highly sparse data, while the bipartite graph learning module can better capture the consensus of the group for medium to low sparse data. Subsequently, with the help of a carefully designed residual gating mechanism, the group representations from these two levels are adaptively balanced, allowing the model to adapt to data with different sparsities. Features are modeled more accurately, which significantly improves the performance on group recommendation tasks and significantly enhances the model's adaptability to data with different sparsities.

Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention.

Description of the drawings

In order to more clearly illustrate the present invention or the technical solutions in the prior art, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are the drawings of the present invention. For some embodiments, those of ordinary skill in the art can also obtain other drawings based on these drawings without exerting creative efforts.

Figure 1 is a flow chart of a group recommendation method based on counterfactual enhanced graph convolution network provided by the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention more clear, the technical solutions in the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention. , not all examples. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts fall within the scope of protection of the present invention. The following examples are used to illustrate the invention but are not intended to limit the scope of the invention.

In the description of the embodiments of the present invention, it should be noted that the terms "center", "longitudinal", "horizontal", "upper", "lower", "front", "back", "left" and "right" The orientations or positional relationships indicated by "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. are based on the orientations or positional relationships shown in the accompanying drawings and are only for the convenience of describing this document. The embodiments and simplified descriptions of the invention do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed and operate in a specific orientation, and therefore cannot be construed as limiting the embodiments of the invention. Furthermore, the terms “first”, “second” and “third” are used for descriptive purposes only and are not to be construed as indicating or implying relative importance.

In the description of the embodiments of the present invention, it should be noted that, unless otherwise clearly stated and limited, the terms "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection or a detachable connection. Or integrated connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium. For those of ordinary skill in the art, the specific meanings of the above terms in the embodiments of the present invention can be understood in specific situations.

In the embodiment of the present invention, unless otherwise expressly provided and limited, the first feature "on" or "below" the second feature may be that the first and second features are in direct contact, or the first and second features are in intermediate contact. Indirect media contact. Furthermore, the terms "above", "above" and "above" the first feature is above the second feature may mean that the first feature is directly above or diagonally above the second feature, or simply means that the first feature is higher in level than the second feature. "Below", "below" and "beneath" the first feature to the second feature may mean that the first feature is directly below or diagonally below the second feature, or simply means that the first feature has a smaller horizontal height than the second feature.

In the description of this specification, reference to the terms "one embodiment," "some embodiments," "an example," "specific examples," or "some examples" or the like means that specific features are described in connection with the embodiment or example. , structures, materials or features are included in at least one embodiment or example of embodiments of the present invention. In this specification, the schematic expressions of the above terms are not necessarily directed to the same embodiment or example. Furthermore, the specific features, structures, materials or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, those skilled in the art may combine and combine different embodiments or examples and features of different embodiments or examples described in this specification unless they are inconsistent with each other.

The embodiment of the present invention is described below with reference to FIG. 1 .

The present invention provides a group recommendation method based on counterfactual enhanced graph convolution network, including:

S1: Preprocess public group data, obtain negative example data, and construct group user hypergraph and group project bipartite graph;

In step S1, the group-user hypergraph is constructed based on group-user relationship data, and the group-item bipartite graph is constructed based on group-item interaction data.

In step S1, the negative example data includes an initial representation set constructed based on the ID embedding vector lookup table. The initial representation set includes an initial group representation, an initial user representation, and an initial item representation.

Further, in this stage, the group recommendation data set is first collected, including user historical interaction data, group historical interaction data and group user relationship data. The goal of this stage is to construct a group user hypergraph based on group user relationship data, and at the same time construct a group item bipartite graph based on group item interaction data; build an embedding vector lookup table for users, groups, and items; randomly select from the data set Sampling negative examples of non-interacting users (groups) to construct a data set for training.

Specifically, it includes: first constructing training data. Since the loss function uses pairwise BPR loss, it is necessary to randomly sample multiple negative examples from uninteracted items for each user (group) and item interaction pair to form this algorithm. Training data required for the invented model.

Secondly, a group user hypergraph and a group project bipartite graph are constructed. In order to efficiently mine high-order information between groups, users and projects, the association corresponding to the group user hypergraph is constructed based on the group user relationships in the data set. Matrix, in which a value of 1 means that the user is included in the group, and a value of 0 means that the user is not included in the group. Based on the group-item interactions in the training data, an adjacency matrix corresponding to the group-item bipartite graph is constructed, where a value of 1 indicates that the group interacts with the item, and a value of 0 indicates that there is no interaction.

Finally, an embedding vector lookup table is constructed, and an ID embedding vector lookup table is constructed based on the number of users, groups, and projects as their respective initial representations.

S2: Convolve the group user hypergraph to mine high-order information among multiple users between groups and users, and obtain user-level group representation; perform perturbation on the group user hypergraph, and perform perturbation on the perturbed The group user hypergraph is convolved to obtain counterfactual group representation;

Furthermore, user-level group representation is member-level group representation. The goal of this stage is to construct a member preference counterfactual hypergraph learning module and use hypergraph convolution operations to extract high-order information between groups and users to obtain Member-level group representation combines hypergraph neural networks with counterfactual learning through the member preference counterfactual hypergraph learning module, aiming to mine the causal relationship between group members and item recommendations. By minimizing the counterfactual loss, the factual prediction score calculated based on the factual and counterfactual group representations is as large as possible than the counterfactual prediction score.

Wherein, in step S2, the expression expressed by the user-level group is:

in, Represented as a user-level group,/> is the total number of convolutional layers of the hypergraph convolutional network,/> is the convolution layer index value of the hypergraph convolution network,/> For the first/> Aggregated information represented by nodes updated by convolutional layers.

Among them, in step S2, the calculation formula for perturbing the group user hypergraph is:

in, is the perturbed group user hypergraph correlation matrix,/> is the group user hypergraph correlation matrix,/> is the indicator function,/> is a trainable mask matrix,/> Threshold specified for indicator function decision entries.

Specifically, it includes: first obtaining the member-level group representation, convolving the constructed group user hypergraph and the initial representation of users and groups through hypergraph convolution to obtain the member-level group representation, and then in order to enhance the expressiveness, the obtained results of each layer are The embedding representations are averaged to obtain the user representation at this level. Similarly, on each convolutional layer, we average the aggregate information of the nodes updated in the previous layer during the message passing process to obtain the group representation at this level.

Secondly, the counterfactual group representation is obtained, and the original hypergraph correlation matrix is perturbed through a trainable mask matrix. Based on the obtained perturbed hypergraph correlation matrix, the same hypergraph convolution operation as above is used to obtain the counterfactual scenario. group representation.

Finally, the counterfactual loss is calculated and optimized based on different group representations and initial item representations in factual and counterfactual scenarios through a three-layer multi-layer perceptron. Calculate factual and counterfactual prediction scores, then calculate the counterfactual loss/> .

S3: Convolve the group-item bipartite graph and the user-level group representation to mine high-order information between groups and items, obtain group-level group representation and group-level item representation, and pass the residual gate A control mechanism weightedly fuses the user-level group representation and the group-level group representation to obtain the final group representation;

Furthermore, this stage mainly uses the group consensus graph learning module and the residual fusion module. The goal of this stage is to construct the group consensus graph learning module and the residual fusion module, and use graph convolution operations to extract the differences between groups and projects. Higher-order information is obtained to obtain group-level group representations, leveraging supervisory signals between group items to enhance model performance when more data becomes available. Then the group representations at different levels are fused into the final group representation through the residual fusion module. By minimizing the group loss, the group's prediction score for positive samples is as much as possible greater than the prediction score for negative samples.

Among them, the expressions of the group-level group representation and group-level item representation obtained in step S3 are:

in, For group-level group representation,/> Represented as a group-level item,/> For the first/> The node representation obtained by the convolutional layer, Average of node embedding representations obtained for all convolutional layers.

Wherein, the expression of the final group representation in step S3 is:

in, is the final group representation,/> Represents the weight of user-level groups,/> is the sigmoid function,/> is the first gating coefficient,/> is the second gating coefficient,/> is the third weight coefficient.

Specifically, it includes: firstly, the constructed group-item bipartite graph and the calculated group representation and the initial representation of the item are used to obtain the group-level group representation through graph convolution, and then the embeddings obtained at each layer are averaged to obtain the final Embed.

Secondly, the calculated member-level group representation and the calculated group-level group representation are used to calculate the weight of each part through the gating mechanism, and the weight of the member-level group representation is Multiply it by itself and add the group-level group representation to its weight/> The result of multiplication is the final group representation.

Finally, the final group representation and group-level item representation are input into the multi-layer perceptron to calculate the prediction score, and then calculate the group loss .

S4: Obtain counterfactual loss based on the counterfactual group representation, user-level group representation and initial item representation, obtain group loss based on the final group representation and the group-level item representation, based on the initial user representation and The initial project indicates that the user loss is obtained, and the hierarchical graph convolution network is trained for multiple rounds according to the counterfactual loss, the group loss and the user loss to obtain a group prediction model;

Wherein, in step S4, the losses used for the multi-round training of the hierarchical graph convolution network include:

in, is the counterfactual loss,/> For group losses,/> Loss for users,/> For the group project training set,/> Data from the training set for the group project,/> Training set for user projects,/> Data from the training set for user projects,/> is the item index value,/> is the sampling negative example corresponding to the item that interacts with the group,/> is the sampling negative example corresponding to the item that interacts with the user,/> Predict the score for the fact,/> is the counterfactual prediction score,/> represents the predicted score for the final group,/> Represents predicted scores for group-level items,/> The user initially indicates the predicted score,/> Predict the score for the initial representation of the project.

S5: Input the group data to be recommended into the prediction model, obtain a prediction score, and obtain a list of recommended items based on the prediction score sorting.

Further, after multiple rounds of joint training, the trained model is obtained, and any group and item representation that needs to be predicted is used as input, and its corresponding prediction score is output, and the group's prediction score for all candidate items is ranked from highest to highest. Take the top K items in low order to obtain the final list of recommended items.

Further, it also includes obtaining the initial representation of the user and the item, and minimizing the user loss so that the user's prediction score for the positive sample is as large as possible greater than the prediction score for the negative sample. Specifically, input the initial representation of the user and item at most Calculate the prediction score in the layer perceptron, and then calculate the user loss and optimize.

In some embodiments, the present invention conducted experiments using two widely used real-world group recommendation data sets and a semi-synthetic group recommendation data set, where the basic statistical properties of the used data sets are shown in Table 1 .

Table 1 Basic statistical characteristics of the data sets used in the embodiments of the present invention

The experiment is divided into two aspects: project recommendation for a group of users, and project recommendation for a single user, which are called group recommendation tasks and user recommendation tasks respectively. In these two experiments, we used two commonly used evaluation indicators: HR (hit rate) and NDCG (normalized loss cumulative gain). The obtained method provided by the present invention is consistent with other methods in the field of group recommendation. The results on each task of the data set are shown in Table 2 and Table 3.

Table 2 Comparison of the first set of experimental results in the data set between the method of the present invention and other methods in the field of group recommendation

Table 3 Comparison of the second set of experimental results in the data set between the method of the present invention and other methods in the field of group recommendation

The experimental results in Table 2 and Table 3 show that our proposed method CF-hGCN performs better than previous methods and has achieved varying degrees of improvement in performance, with the maximum improvement being 12.7%. The above comparison results fully prove that our proposed method achieves excellent results in group recommendation and user recommendation tasks.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that it can still be used Modifications are made to the technical solutions described in the foregoing embodiments, or equivalent substitutions are made to some of the technical features; however, these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (8)

1. A group recommendation method based on counterfactual enhanced graph convolution network, which is characterized by: S1: Preprocess public group data, obtain negative example data, and construct group user hypergraph and group project bipartite graph; S2: Convolve the group user hypergraph to mine high-order information between groups and users, and obtain user-level group representation; perform perturbation on the group user hypergraph, and perform perturbation on the group user hypergraph. The graph is convolved to obtain counterfactual group representation; S3: Convolve the group-item bipartite graph and the user-level group representation to mine high-order information between groups and items, obtain group-level group representation and group-level item representation, and pass the residual gate A control mechanism weightedly fuses the user-level group representation and the group-level group representation to obtain the final group representation; S4: Obtain counterfactual loss based on the counterfactual group representation, user-level group representation and initial item representation, obtain group loss based on the final group representation and the group-level item representation, based on the initial user representation and The initial project indicates that the user loss is obtained, and the hierarchical graph convolution network is trained for multiple rounds according to the counterfactual loss, the group loss and the user loss to obtain a group prediction model; S5: Input the list of items to be recommended into the prediction model, obtain a prediction score, and obtain a list of recommended items based on the ranking of the prediction score. 2. A group recommendation method based on counterfactual enhanced graph convolution network according to claim 1, characterized in that, in step S1, the group user hypergraph is constructed based on the group user relationship data, so The group-item bipartite graph described above is constructed based on the group-item interaction data. 3. A group recommendation method based on counterfactual enhanced graph convolutional network according to claim 1, characterized in that, in step S1, the negative example data includes an initial representation constructed based on an ID embedding vector lookup table. The initial representation set includes an initial group representation, an initial user representation, and an initial item representation. 4. A group recommendation method based on counterfactual enhanced graph convolution network according to claim 1, characterized in that, in step S2, the expression of the user-level group representation is: in, Represented as a user-level group,/> is the total number of convolutional layers of the hypergraph convolutional network,/> is the convolution layer index value of the hypergraph convolution network,/> For the first/> Aggregated information represented by nodes updated by convolutional layers. 5. A group recommendation method based on counterfactual enhanced graph convolution network according to claim 4, characterized in that, in step S2, the calculation formula for perturbing the group user hypergraph is: in, is the perturbed group user hypergraph correlation matrix,/> is the group user hypergraph correlation matrix,/> is the indicator function, is a trainable mask matrix,/> Threshold specified for indicator function decision entries. 6. A group recommendation method based on counterfactual enhanced graph convolution network according to claim 5, characterized in that the expressions of the group-level group representation and the group-level item representation obtained in step S3 for: in, For group-level group representation,/> Represented as a group-level item,/> For the first/> The node representation obtained by the convolutional layer, /> Average of node embedding representations obtained for all convolutional layers. 7. A group recommendation method based on counterfactual enhanced graph convolution network according to claim 6, characterized in that the expression of the final group representation in step S3 is: in, is the final group representation,/> Represents the weight of user-level groups,/> is the sigmoid function,/> is the first gating coefficient,/> is the second gating coefficient,/> is the third weight coefficient. 8. A group recommendation method based on counterfactual enhanced graph convolution network according to claim 7, characterized in that, in step S4, the hierarchical graph convolution network is used for multi-round training. Losses include: in, is the counterfactual loss,/> For group losses,/> Loss for users,/> For the group project training set,/> Data from the training set for the group project,/> Training set for user projects,/> Data from the training set for user projects,/> is the item index value,/> is the sampling negative example corresponding to the item that interacts with the group,/> is the sampling negative example corresponding to the item that interacts with the user,/> Predict the score for the fact,/> is the counterfactual prediction score,/> represents the predicted score for the final group,/> Represents predicted scores for group-level items,/> The user initially indicates the predicted score,/> Predict the score for the initial representation of the project.