Multi-behavior fused graph convolutional network recommendation method and device

Abstract

The invention discloses a multi-behavior fused graph convolutional network recommendation method and device. The method comprises the following steps: acquiring a preset user embedding vector, a preset project embedding vector and a preset behavior embedding vector; on the basis of a message passing architecture of a graph convolutional neural network, capturing graph structure information of heterogeneous interaction of user items and multi-behavior cooperative signals, and calculating to obtain user embedded representation, item embedded representation and behavior embedded representation after propagation; performing behavior interaction of the same type on any two items to obtain common features and association between the any two items, and calculating to obtain item correlation embedding representation; and according to the user embedded representation, the project embedded representation, the behavior embedded representation and the project correlation embedded representation, predicting the future behavior of the user on the project. According to the technical scheme, the recommendation accuracy and robustness are remarkably improved, and the method is suitable for various complex scenes.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent algorithms, and in particular to a graph convolutional network recommendation method and device integrating multiple behaviors. Background Art

[0002] Recommendation systems have become an important part of online platforms, among which collaborative filtering is a widely used method that provides users with more accurate recommendations by learning users' interests and estimating users' preferences from user behavior data. Collaborative filtering is mainly divided into memory-based collaborative filtering and model-based collaborative filtering. Memory-based collaborative filtering is further divided into user-based collaborative filtering and item-based collaborative filtering according to different entities. User-based collaborative filtering mainly uses the similarity matrix of the target user to find other users with similar interests and preferences, and recommends items that these user groups like but the target user has not yet known to the target user. This method helps to explore the user's potential attributes. Item-based collaborative filtering mainly recommends items similar to the user's historical preferences based on the similarity matrix of the items. Model-based collaborative filtering uses the user's historical preference data to train a mathematical model to predict the ratings of uninteracted items and make recommendations based on the ratings, which can avoid the problems of insufficient memory and high computational complexity that may occur in memory-based methods when the data volume is large. Common models include collaborative filtering algorithms based on clustering, Bayesian, probability, maximum entropy, and matrix decomposition. Although these methods have achieved good results, as shallow models, they are designed to be scalable and have difficulty adapting to the increasingly complex and varied scale and form of data on the Internet.

[0003] Content-based recommendation systems recommend similar items based on the attribute characteristics of the items themselves, such as the category, director, and actor of the movie, based on the user's historical preferences. The similarity between the feature vectors of item attributes is usually expressed by cosine similarity. This method does not require a large number of rating records, so it is less affected by data sparsity and cold start problems. However, due to the increasing attention paid to personal privacy by users, it has become difficult to collect user preference information. At the same time, this traditional shallow model is not highly scalable in design, and the effective feature extraction of items is limited, which also limits the quality of recommendations. Deep learning-based recommendation systems use relevant information of different entities (such as users, items, etc.) as input, learn the feature representation of entities from samples through deep learning models, and use these feature representations to generate recommendations. Among them, the classic recommendation model based on multi-layer perceptrons sets multiple hidden layers between the input layer and the output layer, and introduces activation functions to enhance the nonlinear fitting ability of the model to improve accuracy. Convolutional neural networks (CNNs) can map features and user preferences to the same latent space due to their ability to extract global and local features, thereby improving the accuracy of the model. Some studies apply CNN to recommendation tasks where the raw data is structured multimedia data (such as images, text, or audio). In sequence data, there is a temporal dependency between data. Recurrent neural networks (RNNs) mine the temporal dependency information of sequence data through loops and memories, effectively modeling users' dynamic preferences. However, these models are very effective in processing regular Euclidean data (such as images and text), but cannot be applied to non-Euclidean space data such as graph structures. With the increase of graph structure data, such as social networks (users are nodes, and relationships between users are edges), molecular graphs in biochemistry (atoms are nodes, chemical bonds are edges), and transportation networks (locations are nodes, and driving paths are edges), a new type of artificial neural network has emerged, called graph neural network. This includes graph convolutional neural network (GCN), graph attention network (GAT), graph autoencoder (GAE), etc. However, these models mainly focus on a single type of user-item interaction behavior, such as purchase behavior in e-commerce platforms, which may lead to cold start or data sparsity problems in real applications. In real life, user behavior is essentially multi-type, such as like, favorite, add to cart, etc. How to properly solve the above problems has become an urgent issue to be solved in the industry. Summary of the invention

[0004] The present invention provides a graph convolutional network recommendation method, device, equipment and storage medium that integrates multiple behaviors, which are used to improve the accuracy and robustness of the recommendation system and make it suitable for various complex scenarios.

[0005] According to a first aspect of the present invention, a graph convolutional network recommendation method integrating multiple behaviors is provided, and the graph convolutional network recommendation method integrating multiple behaviors includes:

[0006] Obtain the preset user embedding vector, item embedding vector, and behavior embedding vector as input features of the graph convolutional network;

[0007] Based on the message passing architecture of graph convolutional neural network, the user-item embedding propagation layer with behavior awareness captures the graph structure information of heterogeneous user-item interactions and the collaborative signals of multiple behaviors, and calculates the user embedding representation, item embedding representation and behavior embedding representation after propagation;

[0008] Perform the same type of behavioral interaction on any two items, obtain the common features and associations between the two items, and calculate the item correlation embedding representation;

[0009] The user's future behavior on the project is predicted based on the user embedding representation, the project embedding representation, the behavior embedding representation and the project relevance embedding representation.

[0010] In one embodiment, the message passing architecture based on graph convolutional neural network captures the graph structure information of heterogeneous user-project interactions and the collaborative signals of multiple behaviors, including:

[0011] Based on the message passing architecture of graph convolutional neural network, a behavior-aware user-item embedding propagation layer is established;

[0012] The user-item embedding propagation layer captures the graph structure information of user heterogeneous interactions and the collaborative signals of multiple behaviors.

[0013] In one embodiment, the user embedding representation includes:

[0014] Rewrite the convolutional layer propagation formula of the graph convolutional neural network, embed and combine the relationship between adjacent project nodes to model the interaction behavior between users and projects;

[0015] Different interactive behaviors of the user are assigned respective propagation weights, and the propagation weights integrate the intensity and quantity of the behaviors.

[0016] In one embodiment, obtaining a preset user embedding vector, an item embedding vector, and a behavior embedding vector as input features of a graph convolutional network includes:

[0017] The user embedding matrix U is composed of n user embedding vectors, the item embedding matrix V is composed of m item embedding vectors, and the behavior embedding matrix E is composed of t behavior embedding vectors, where the number of behavior types is determined by the type of interaction between users and items;

[0018] The user embedding matrix U, the item embedding matrix V and the behavior embedding matrix E are used as input features of the graph convolutional network.

[0019] In one embodiment, performing the same type of behavior interaction on any two items to obtain common features and associations between the any two items includes:

[0020] Performing the same type of behavioral interaction on any two items to obtain common features and associations between the two items;

[0021] It captures the similarity or correlation of different behaviors in items and aggregates the features of items of the same type into the item embedding vector.

[0022] In one embodiment, predicting the user's future behavior on the project includes:

[0023] Through propagation of a preset number of layers, multiple embedding representations of users, items, and behaviors are obtained and aggregated using preset weights;

[0024] Calculate the prediction score of any user performing a certain action on any item. The prediction score not only captures the high-order collaborative signals of users and items under multiple actions, but also represents the semantics of the actions.

[0025] The prediction score is optimized by minimizing a preset loss function.

[0026] According to a second aspect of the present invention, a graph convolutional network recommendation device integrating multiple behaviors is provided, comprising:

[0027] An acquisition module is used to obtain preset user embedding vectors, item embedding vectors, and behavior embedding vectors as input features of the graph convolutional network;

[0028] The first computing module is used for the message passing architecture based on graph convolutional neural network. Through the behavior-aware user-item embedding propagation layer, it captures the graph structure information of user-item heterogeneous interactions and the collaborative signals of multiple behaviors, and calculates the propagated user embedding representation, item embedding representation and behavior embedding representation;

[0029] The second calculation module is used to perform the same type of behavioral interaction on any two items and calculate the item correlation embedding representation;

[0030] A prediction module is used to predict the user's future behavior on the project based on the user embedding representation, the project embedding representation, the behavior embedding representation and the project relevance embedding representation.

[0031] According to a third aspect of the present invention, there is provided an electronic device, the electronic device comprising: a communication interface, a processor, and a memory;

[0032] Among them, the memory is used to store program instructions, and when the program instructions are executed by the processor that is communicatively connected to the memory through the communication interface, any of the above-mentioned graph convolutional network recommendation methods that integrate multiple behaviors is implemented.

[0033] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a computer (e.g., a processor in a computer), any of the above-mentioned graph convolutional network recommendation methods that integrate multiple behaviors is implemented.

[0034] In summary, the present invention provides a graph convolutional network recommendation method and device integrating multiple behaviors, the method comprising: obtaining preset user embedding vectors, project embedding vectors and behavior embedding vectors as input features of the graph convolutional network; based on the message passing architecture of the graph convolutional neural network, through the user-project embedding propagation layer with behavior awareness, capturing the graph structure information of the heterogeneous interaction of user projects and the collaborative signals of multiple behaviors, and calculating the propagated user embedding representation, project embedding representation and behavior embedding representation; performing the same type of behavioral interaction on any two projects, and calculating the project correlation embedding representation; predicting the user's future behavior on the project according to the user embedding representation, project embedding representation, behavior embedding representation and project correlation embedding representation. The technical solution of the present application learns the collaborative signals between heterogeneous data through the powerful learning ability of the graph neural network for the graph structure; distinguishes the importance of different behaviors by setting propagation weights; and enhances the learning ability of project embedding by extracting the correlation between projects, and finally achieves the technical effect of improving the recommendation accuracy.

[0035] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings.

[0036] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0038] Figure 1A flowchart of a graph convolutional network recommendation method integrating multiple behaviors provided by an embodiment of the present invention;

[0039] Figure 2 A flowchart of step S12 of a graph convolutional network recommendation method integrating multiple behaviors provided by an embodiment of the present invention;

[0040] Figure 3 A flowchart of step S11 of a graph convolutional network recommendation method integrating multiple behaviors provided by an embodiment of the present invention;

[0041] Figure 4 A flowchart of step S13 of a graph convolutional network recommendation method integrating multiple behaviors provided by an embodiment of the present invention;

[0042] Figure 5 A flowchart of step S14 of a graph convolutional network recommendation method integrating multiple behaviors provided by an embodiment of the present invention;

[0043] Figure 6 A structural diagram of a graph convolutional network recommendation device integrating multiple behaviors provided by an embodiment of the present invention;

[0044] Figure 7 A structural diagram of an electronic device provided by an embodiment of the present invention;

[0045] Figure 8 An overall framework diagram of a graph convolutional network recommendation method integrating multiple behaviors provided by an embodiment of the present invention;

[0046] Fig. 9 A schematic diagram of a message passing architecture of a graph convolutional network recommendation method integrating multiple behaviors provided by an embodiment of the present invention;

[0047] Fig.10 A schematic diagram of item relevance extraction of a graph convolutional network recommendation method integrating multiple behaviors provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The features and exemplary embodiments of various aspects of the present application will be described in detail below. In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain the present application, rather than to limit the present application. For those skilled in the art, the present application can be implemented without the need for some of these specific details. The following description of the embodiments is only to provide a better understanding of the present application by illustrating the examples of the present application.

[0049] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "include..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0050] like Figure 1 As shown, the present invention provides a graph convolutional network recommendation method integrating multiple behaviors, and the graph convolutional network recommendation method integrating multiple behaviors includes:

[0051] In step S11, a preset user embedding vector, item embedding vector and behavior embedding vector are obtained as input features of the graph convolutional network;

[0052] In step S12, based on the message passing architecture of graph convolutional neural network, the graph structure information of heterogeneous user-item interactions and the collaborative signals of multiple behaviors are captured through the behavior-aware user-item embedding propagation layer, and the propagated user embedding representation, item embedding representation and behavior embedding representation are calculated;

[0053] In step S13, the same type of behavior interaction is performed on any two items to obtain common features and associations between the two items, and an embedding representation of the item correlation is calculated;

[0054] In step S14, the user's future behavior on the project is predicted based on the user embedding representation, the project embedding representation, the behavior embedding representation and the project relevance embedding representation.

[0055] In one embodiment, users, items (such as commodities, movies, etc.) and user behaviors on items (such as purchases, likes, comments, etc.) are converted into numerical vectors, which are called embedding vectors. User embedding vectors may contain basic information of users, historical behavior patterns, etc.; item embedding vectors may contain attribute features of items; and behavior embedding vectors represent specific behavior types of users on items. The embedding vectors are used as input features of the graph convolutional network to provide basic data for subsequent model training and prediction. The graph convolutional network uses its message passing mechanism to update the embedding vectors of users, items, and behaviors. Message passing refers to the network passing information through the edges in the graph structure (i.e., the interactive relationship between users and items), thereby updating the representation of nodes (i.e., users and items). The behavior-aware embedding propagation layer means that when updating the embedding vector, the network not only considers the connections between users and items, but also the types of behaviors on these connections. The model can capture heterogeneous interaction information between users and items (i.e., different types of behaviors) and synergistic signals between multiple behaviors (i.e., the mutual influence between different behaviors), and can obtain updated user embedding representations, item embedding representations, and behavior embedding representations. By analyzing the pattern of users performing the same type of behavior on two different items, the system can infer the similarity or correlation between the two items. For example, if many users have made purchases on item A and item B, then it can be considered that item A and item B are related to some extent. This correlation is expressed in the form of an embedding vector, called the item relevance embedding representation. Predicting a user's possible future behavior through an embedding vector involves a classification model that determines whether the user is interested in a certain item based on the user's embedding representation, or predicts the specific type of behavior the user may take based on the behavior embedding representation. The item relevance embedding representation can help the model better understand the similarities between items, thereby providing users with more accurate recommendations. The complex relationships between users, items, and behaviors are handled through a graph convolutional network, and the user's future behavior is predicted by capturing and learning the patterns of these relationships, thereby providing personalized recommendations.

[0056] The model of the present invention mainly includes four main components: shared embedding module, user-item message transmission module, item relevance extraction module and the final prediction module. The overall framework diagram is as follows Figure 8 As shown in the figure, the specific functions of each module are as follows:

[0057] 1. Shared Embed Module

[0058] Using Embedding Vectors to describe user i, item j, and relationship k, and d is the embedding dimension. The user embedding vector and item embedding vector can be represented by embedding matrices U and V respectively. E is the behavior type embedding matrix corresponding to the interaction type between users and items. The formula is as follows:

[0059]

[0060] Among them, n, m, and T are their corresponding quantities respectively. The embedded representations in the matrices U, V, and E can be used as initialization features of users, items, and behavior types, and as input features of users, items, and relationships in the graph neural network architecture.

[0061] 2. User project messaging module

[0062] Based on the GCN message passing architecture, a behavior-aware user-item embedding propagation layer is established to capture the graph structure information of heterogeneous user-item interactions and high-order collaborative signals under multiple behaviors. In this model, the embedding of each node (user and item) is modeled by accumulating the incoming information of all its neighboring nodes, and the behavioral relationship between nodes is incorporated into the recommendation to strengthen the embedding learning and thus improve the prediction effect, such as Fig. 9 shown.

[0063] Rewrite the GCN convolutional layer propagation formula to calculate the relationship between adjacent project nodes Embedding combinations are performed to model the interaction between users and items. At the same time, different behaviors have different impacts on users, and different behaviors have different strengths and interaction sparsity. For example, the target behavior has the greatest importance among all behaviors, but the number of target behavior interactions between users and items is usually small. Using other auxiliary behaviors can help learn the target behavior. Based on this, a propagation weight α is assigned to the user. This parameter combines the strength and number of behaviors. The formula is as follows:

[0064]

[0065] For each behavior, It is learned through the model and is the intensity weight of behavior k. is the number of interactions performed by user u using behavior k. The total number of all types of behaviors of user u is recorded as N r is the set of all user behaviors. The final weight α of user u under behavior k is calculated taking into account both the importance and sparsity of the behavior.

[0066] Then, the neighbor nodes v of user u and the relationship k between them are combined, and the Hadamard product is used as the combination operation method. (l) It is a weight parameter matrix specific to the graph convolution layer from layer l to layer l+1. u and N v are the neighbor node sets of user u and project v respectively. As a symmetric normalization, we avoid the embedding scale from increasing with the number of convolutional layers. The activation function δ is LeakyReLU, and we get is the embedded representation of user u obtained after the l-th layer of convolutional propagation. It is also a relation parameter matrix specific to the lth layer to the l+1th layer, mapping all interaction relations and nodes to the same embedding space.

[0067] The model of the present invention assigns propagation weights to different behaviors in the former, because it is considered that the items interacted with by different behaviors can reflect the preference characteristics of different users to a certain extent. However, the characteristics of the items themselves are static, so there is no need to set weight parameters for the items.

[0068]

[0069] Among them, the neighbor node u of item v and the relationship k between them are combined, and the relationship embedding is included in the formula. The final result is is the embedding representation of item v obtained after the l-th layer of convolutional propagation.

[0070] 3. Project correlation extraction module

[0071] The same type of behavioral interactions on different items can reflect that there should be some common features and associations between items, and the relationship between items may affect user behavior. Capturing the similarities or associations between items in different behaviors can enhance the learning ability of item embedding, such as Fig.10 shown.

[0072]

[0073] in, is a set of items that interact with the user under behavior k, just like item v. It is a parameter matrix of row k, which helps to aggregate information at the l-1th layer, and the aggregation function is a simple average function. is the initial embedding vector of the corresponding item v, and the embedding of item j with the same behavior k as item v is aggregated Get the next layer of embedding In this way, the features of other items that have the same type of behavior as item v are aggregated into v.

[0074] 4. Prediction Module

[0075] After L layers of propagation, multiple embedding representations of user u, item v, and relationship k are obtained. The embedding representations obtained at different layers represent messages transmitted from neighbors of different orders. Further aggregation is performed to obtain the final corresponding representation. Each embedding layer is set with a uniform weight. The formula is as follows:

[0076]

[0077] Establish the probability score of the target user u performing the kth action on the target item v, and calculate the correlation score between the target item v and the items that the user u has interacted with under the specific action k. The formula is as follows:

[0078]

[0079] Here, diag(·) represents a diagonal matrix, and the diagonal elements are correspondingly equal to Represents the set of items for user u under behavior k. Combining the above two formulas with the hyperparameter λ as the final prediction score, the final result not only captures the high-order collaborative signals of users and items under multiple behaviors, but also considers the behavior semantics:

[0080] y(k) u，v =λ·y1(k) u，v +(1-λ)·y2(k) u，v

[0081] The goal of model optimization is to minimize the loss function of the following formula:

[0082]

[0083] Where I is the number of trained users, S is the number of positive and negative logarithms for each user. For each user, randomly select S positive interaction items v for the target behavior. p1 、v p2 ...v ps and the same number of negative interaction items v that have no interaction with the user n1 、v n2 ...v ns The first term is the pairwise loss of positive and negative pairs, the second term θ represents the set of all trainable parameters, and λ' is the coefficient of the regularization term. During the training phase, the Adam algorithm is used to optimize the above defined objectives.

[0084] The technical solution in this embodiment uses the powerful learning ability of graph neural networks for graph structures to learn collaborative signals between heterogeneous data; distinguishes the importance of different behaviors by setting propagation weights; and enhances the learning ability of project embedding by extracting the correlation between projects, ultimately achieving the technical effect of improving recommendation accuracy.

[0085] In one embodiment, Figure 2 As shown, S12 includes the following steps S21-S24:

[0086] In step S21, a user-item embedding propagation layer with behavior awareness is established based on the message passing architecture of the graph convolutional neural network;

[0087] In step S22, the graph structure information of user heterogeneous interactions and the collaborative signals of multiple behaviors are captured through the user-item embedding propagation layer;

[0088] In step S23, the convolutional layer propagation formula of the graph convolutional neural network is rewritten to embed and combine the relationship between adjacent project nodes to model the interaction behavior between users and projects;

[0089] In step S24, respective propagation weights are allocated to different interactive behaviors of the user, and the propagation weights integrate the intensity and quantity of the behaviors.

[0090] In one embodiment,

[0091] Based on the message passing architecture of the Graph Neural Network (GNN), the information about the interaction between users and items is extracted through the behavior embedding propagation layer. Since the interaction between users and items may include multiple behaviors such as clicks, purchases, and favorites, the interaction characteristics of these interactions can be captured through the behavior collection layer, so that the harmonious signals of multiple behaviors can be extracted from the graph structure information, allowing the model to have a more comprehensive understanding of the relationship between users and items. By integrating user behavior data through the design of the graph neural network, the model can have a more accurate understanding of user preferences.

[0092] In the message forwarding architecture of the graph neural network, by setting up a user embedding propagation layer for behavior capture, multiple interaction modes between users and items can be separated and identified. This embedding propagation layer is logically designed to provide a deeper structure for different user behaviors (such as clicks, browsing, purchases, etc.), so that these behaviors can be learned as independent features in the model, establishing the ability to distinguish behaviors and providing support for the generation of embedding representations of each behavior in subsequent steps.

[0093] The user-item embedding propagation layer captures the graph structure information of different behaviors between users and items, as well as the good signals between these behaviors. A clearer expression of the captured graph structure information and behavior signal relationship between users and items can be provided in the model, so that the model can not only understand a single behavior, but also identify the consistency between multiple behaviors, thereby more comprehensively characterizing the user's preferences and the value of the project. Logically, this layer makes the user embedding representation and the item embedding representation more expressive by well integrating the information of different behaviors.

[0094] By rewriting the topological layer propagation formula of the graph neural network, the model can more accurately model the interaction between users and projects, better combine the embedding relationship between user and project nodes, and better capture the interaction between project nodes in the model. This enables the graph neural network to establish a more expressive relationship model between users and project nodes.

[0095] Assigning independent propagation weights to different user behaviors (such as clicks and purchases) enables the model to more effectively mitigate the intensity and frequency of behaviors. Propagation weights can integrate information about the intensity and quantity of behaviors, helping the model prioritize user preferences. Through personalized weight mechanism allocation, the model is provided with a means to distinguish user behaviors, and can more accurately focus on the degree of preference of users in different behaviors, providing support for the final prediction and recommendation.

[0096] Through the message passing architecture of graph neural networks, a behavior-aware embedded propagation layer is established to capture the conversational interaction characteristics between users and items. The harmonious signals of different behaviors are further captured through the mechanism of behavioral differences. The inner layer propagation formula is rewritten to better combine and express the relationship between users and items. Independent propagation weights are assigned to different behaviors to accurately shape users' behavioral preferences, thereby improving the accuracy of recommendations.

[0097] In one embodiment, Figure 3 As shown, S11 includes the following steps S31-S32:

[0098] In step S31, the n user embedding vectors form a user embedding matrix U, the m item embedding vectors form a project embedding matrix V, and the t behavior embedding vectors form a behavior embedding matrix E, where the number of behavior types is determined by the interaction type between the user and the item;

[0099] In step S32, the user embedding matrix U, the item embedding matrix V and the behavior embedding matrix E are used as input features of the graph convolutional network.

[0100] In one embodiment, the input features of the graph embedding network are constructed, specifically by obtaining preset user embedding tags, project embedding tags, and behavior embedding tags, constructing a corresponding embedding matrix, and providing it as an input feature to the graph embedding network. The user embedding tag is to generate or obtain an embedding vector in advance for each user, representing the characteristics or attributes of the user. The project embedding tag is to generate or obtain an embedding vector in advance for each project, representing the characteristics or attributes of the project. The behavior embedding tag is to generate or obtain an embedding vector in advance for each interaction behavior between the user and the project, representing different behavior types. The preset embedding tags provide an initial feature representation for the graph embedding network, which facilitates the subsequent modeling of the complex relationship between users, projects, and behaviors.

[0101] The user embedding matrix consists of n user embedding vectors. Each row represents the embedding vector of a user, and the matrix size is n*d, where d is the dimension of the embedding vector. The item embedding matrix V consists of m item embedding vectors. Each row represents the embedding vector of an item, and the matrix size is m*d. The behavior embedding matrix E consists of t behavior embedding vectors. Each row represents the embedding vector of a behavior type, and the matrix size is t*d.

[0102] Determine the number of behavior types t. The number of behavior types is determined by the type of interaction between the user and the item. If there are three types of interaction between the user and the item: "click", "collect", and "buy", then t = 3. The number of behavior types can be clarified to help the model accurately capture the impact of different interaction behaviors on the relationship between users and items.

[0103] The embedding matrix is ​​used as the input feature of the graph embedding network, where the user embedding matrix U provides the initial representation of user features; the item embedding matrix V provides the initial representation of item features; and the behavior embedding matrix E provides the initial representation of different behavior types. Inputting these embedding matrices into the graph embedding network enables the model to utilize graph structure information and learn the complex interactive relationships between users, items, and behaviors.

[0104] Get the preset user embedding tags, item embedding tags, and behavior embedding tags. Construct the user embedding matrix U, item embedding matrix V, and behavior embedding matrix E. Provide these embedding matrices as input features to the graph embedding network. The graph embedding network uses these input features and combines graph structure information to learn and model the complex interactive relationships between users, items, and behaviors. Through the above process, the model can capture the interaction patterns between users and items under different behaviors, improve the understanding of user interests and preferences, and thus improve the accuracy of recommendations or predictions.

[0105] This text describes the process of using preset user, project and behavior embedding tags to construct corresponding embedding matrices U, V and E, and provide them as input features to the graph embedding network. Integrate multi-source data and uniformly represent user features, project features and interactive behavior features as an embedding matrix. Model complex relationships, and use graph structural information through graph embedding networks to deeply learn the multi-dimensional interactive relationships between users, projects and behaviors. Improve model performance by accurately capturing the impact of different interactive behaviors, enhance the model's understanding of user preferences, and improve the effectiveness of recommendation systems or other tasks. The technical solution in this embodiment can provide rich input features and structural information, enabling it to more comprehensively understand complex patterns and relationships in the data.

[0106] In one embodiment, Figure 4 As shown, S13 includes the following steps S41-S42:

[0107] In step S41, the same type of behavior interaction is performed on any two items to obtain common features and associations between the any two items;

[0108] In step S42, the similarity or correlation of different behaviors in items is captured, and the features of items of the same type are aggregated into the item embedding vector.

[0109] In one embodiment, the same type of behavioral interactions are performed between any two items, and these interactions are used to capture the common features and correlations between the items, thereby aggregating the same type of item features into the item embedding vector.

[0110] Perform the same type of behavioral interaction on any two items to obtain the common features and associations between the two items. Select any two items in the item set, and there are the same type of user behavioral interactions on these two items. For example, the user has performed the "purchase" behavior on both items. Based on the same behavior of the user on these two items, it can be inferred that the two items may have certain common features or associations, such as the same category, similar functions, or meet similar user needs. The same type of behavior performed by users on different items reflects the similarity or association of these items to a certain extent. User behavior data provides a source of information about the association between items. By analyzing these interactions, common features between items, such as themes, styles, uses, etc., can be extracted, which helps to understand the similarity of items in the feature space. Using the same type of behavioral interactions, the degree of association between items, such as co-occurrence frequency, association metrics, etc., can be calculated, providing a basis for subsequent model learning.

[0111] Capture the similarity or correlation of different behaviors in the project, and aggregate the features of the same type of projects into the project embedding vector. In addition to the same type of behavior, different behaviors (such as clicks, favorites, and purchases) may also have similarities or correlations in the project, which need to be considered together. Aggregate the information of projects with the same type of features and integrate them into the project embedding vector to enhance the expressiveness of project features. The correlation between behaviors, different types of behaviors may be associated, for example, "click" may be the preceding behavior of "purchase". Capturing these correlations helps understand the user's behavior path and preferences on the project. Comprehensively analyzing the features of different behaviors in the project, we can obtain the performance of the project in different dimensions. By aggregating the features of the same type of projects into the project embedding vector, the embedding vector contains not only the information of a single project, but also the features of similar projects, which improves the richness of the embedding representation. A more comprehensive project embedding vector helps the model to more accurately predict user preferences and behaviors in the task and improve the accuracy of recommendations.

[0112] Perform the same type of behavioral interactions on any two items, and use these interactive behaviors to discover common features and correlations between items. Capture the similarities or correlations of different behaviors in items, analyze the user's interaction with items under different behaviors, and understand the correlation between behaviors. Aggregate the features of items of the same type into the item embedding vector to enrich the feature representation of the items. Through the above steps, the model can have a deeper understanding of the correlation between items and the user's behavior patterns, thereby improving the accuracy of predicting user interests in the recommendation system.

[0113] By analyzing the similarities and differences of users on different items, the model can capture more subtle correlations between items and provide a basis for calculating similarities between items. Aggregating the features of items of the same type into the embedding vector enhances the comprehensiveness of item representation and provides the model with a richer source of information when processing item features. There may be synergistic or transformation relationships between different types of behavior. Capturing these relationships helps the model understand the user's behavior path and improves its ability to predict future user behavior.

[0114] On e-commerce platforms, by analyzing users' purchase and browsing behaviors on different products, the correlation between products is captured, and the features of similar products are aggregated into product embedding vectors to improve the accuracy of recommendations. In social media, users' likes and comments on posts and pictures can be used to capture the correlation between content and help recommend related content.

[0115] The user's behavioral interactions on different items are used to capture the correlation between items, thereby enriching the embedded representation of the items. By performing the same type of behavioral interactions on any two items, the common features and associations of the items are obtained; at the same time, the similarities or associations of different behaviors in the items are captured, and the features of the same type of items are aggregated into the item embedding vector. This process enhances the model's understanding of item features and user behavior, and helps improve the model's performance in prediction and recommendation tasks.

[0116] In one embodiment, Figure 5 As shown, S14 includes the following steps S51-S53:

[0117] In step S51, multiple embedding representations of users, items, and behaviors are obtained through propagation of a preset number of layers, and aggregated through preset weights;

[0118] In step S52, a prediction score of any user performing a certain action on any item is calculated, wherein the prediction score not only captures the high-order collaborative signals of users and items under multiple actions, but also represents the semantics of the actions;

[0119] In step S53, the prediction score is optimized by minimizing a preset loss function.

[0120] In one embodiment, the process of predicting a user's future execution of a certain behavior on a project is to obtain multiple embedded representations of users, projects and behaviors through propagation of a preset number of layers, and to aggregate them through preset weights. The propagation of the preset number of layers uses the preset number of layers to propagate information layer by layer in the model to obtain multiple embedded representations of users, projects and behaviors. The preset weight aggregation is to weighted aggregate the embedded representations obtained at each layer according to the preset weights to generate the final embedded representation. The multi-layer propagation mechanism is that the model (such as a graph convolutional network) updates the embedded representation of the node layer by layer through the message passing mechanism at the preset number of layers. The embedded representation is obtained in that at each layer, the embedded representation of the user, project and behavior is updated in combination with the information of the neighboring nodes to capture higher-order interactive relationships. High-order feature capture is that through multi-layer propagation, the high-order association between users and projects can be captured, that is, not limited to direct interaction, but also indirect association. The preset weights may set different values ​​according to the number of layers, reflecting the importance of the embedded representations of each layer. The embedded representations of each layer are weighted summed or other aggregation operations are performed according to the preset weights to obtain a comprehensive embedded representation. The aggregated embedding representation fuses information at different levels and has richer features, which facilitates subsequent predictions.

[0121] Calculate the prediction score of any user performing a certain action on any item. The prediction score not only captures the high-order synergy signals of users and items under multiple behaviors, but also represents the semantics of the behavior. Calculate the probability score of a user performing a certain action on an item using the aggregated user, item, and behavior embedding representations. The prediction score reflects the high-order synergy between users and items under multiple behaviors. The prediction score also reflects the semantic information of a specific behavior, and the impact of different behavior types on the prediction score is fully considered. Fusion of embedding representations, fusing the user embedding representation, item embedding representation, and behavior embedding representation, possibly through dot products, concatenation, or other methods. Use a scoring function (such as a neural network or a simple similarity calculation) to calculate the prediction score to measure the likelihood of a user performing a certain action on an item. Considering the interaction between users and items under different behaviors improves the depth of understanding of user interests. Because the embedding representation is propagated through multiple layers, it captures high-order user-item associations, not just direct interactions. The behavior embedding representation contains semantic information of the behavior, enabling the model to distinguish different types of behaviors (such as clicks, purchases). By fusing behavior semantics, the model can give more accurate prediction scores for different behavior types.

[0122] Use the preset loss function to measure the difference between the predicted score and the actual situation. Through the optimization algorithm (such as gradient descent), minimize the loss function and update the model parameters. After multiple iterations, the predicted score gradually approaches the true value, improving the prediction accuracy of the model. The preset loss function may choose cross entropy loss, mean square error, ranking loss, etc., depending on the task requirements. The measurement standard is used for the loss function to measure the difference between the model prediction and the true label to guide the learning direction of the model. Using the optimization algorithm, update the parameters of the model (such as embedding representation, weight, etc.) according to the gradient information of the loss function. Through multiple iterations, the model parameters are continuously optimized so that the loss function value is gradually reduced. By minimizing the loss function, the model's prediction score more accurately reflects the user's true preferences and behavioral tendencies. The optimized model can also maintain good prediction performance on unknown data.

[0123] Through propagation of a preset number of layers, multiple embedding representations of users, items, and behaviors are obtained. Using the preset weights, these embedding representations are aggregated to obtain a comprehensive feature representation. Using the aggregated embedding representation, the prediction score of a user performing a certain behavior on an item is calculated. The prediction score captures both high-order collaborative signals and semantic information of the behavior. By minimizing the preset loss function, the model parameters are optimized to make the prediction score more accurate. The optimization process improves the prediction ability and generalization performance of the model. Multi-layer propagation allows the model to capture higher-order relationships between users and items, beyond the limitations of first-order neighbors. Different user behaviors have different meanings, for example, "click" may indicate interest, while "buy" indicates a strong intention to buy. Considering the semantics of behavior in prediction helps the model reflect the user's true intention more accurately. The loss function guides the learning direction of the model. By minimizing the loss, the model continuously adjusts its own parameters to improve the prediction accuracy. It can predict users' purchase, browsing, collection and other behaviors of goods on e-commerce platforms, and improve the personalization of recommendations. It can also optimize content recommendation strategies and improve user stickiness by predicting users' likes, sharing, and comments on content.

[0124] Through propagation of a preset number of layers and weight aggregation, embedded representations of users, items, and behaviors containing high-order information are obtained; using these embedded representations, prediction scores that can reflect multi-behavior collaborative signals and behavioral semantics are calculated; finally, by minimizing the preset loss function, the model is optimized to improve the accuracy of the prediction and the generalization ability of the model.

[0125] In one embodiment, Figure 6 FIG. 1 is a block diagram of a graph convolutional network recommendation device integrating multiple behaviors according to an exemplary embodiment. Figure 6As shown, the graph convolutional network recommendation device integrating multiple behaviors includes an acquisition module 61, a first calculation module 62, a second calculation module 63 and a prediction module 64.

[0126] The acquisition module 61 is used to acquire a preset user embedding vector, an item embedding vector and a behavior embedding vector as input features of the graph convolutional network;

[0127] The first computing module 62 is used for capturing the graph structure information of heterogeneous user-item interactions and the collaborative signals of multiple behaviors through a user-item embedding propagation layer with behavior awareness based on a message passing architecture of a graph convolutional neural network, and calculating the propagated user embedding representation, item embedding representation and behavior embedding representation;

[0128] The second calculation module 63 is used to perform the same type of behavior interaction on any two items and calculate the item correlation embedding representation;

[0129] The prediction module 64 is used to predict the user's future behavior on the project based on the user embedding representation, the project embedding representation, the behavior embedding representation and the project relevance embedding representation.

[0130] The acquisition module 61, the first calculation module 62, the second calculation module 63 and the prediction module 64 included in the block diagram of the graph convolutional network recommendation device integrating multiple behaviors are controlled to execute the graph convolutional network recommendation method integrating multiple behaviors described in any of the above embodiments.

[0131] like Figure 7 As shown, the present invention provides an electronic device 700, which includes: a communication interface, a processor 701, and a memory 702;

[0132] Among them, the memory 702 is used to store program instructions, and when the program instructions are executed by the processor 701 that is communicatively connected to the memory 702 through the communication interface, the preset user embedding vector, project embedding vector and behavior embedding vector are obtained as input features of the graph convolutional network; based on the message passing architecture of the graph convolutional neural network, the graph structure information of the heterogeneous interaction between users and projects and the collaborative signals of multiple behaviors are captured through the user-project embedding propagation layer with behavior awareness, and the propagated user embedding representation, project embedding representation and behavior embedding representation are calculated; the same type of behavior interaction is performed on any two projects, the common features and associations between the any two projects are obtained, and the project correlation embedding representation is calculated; based on the user embedding representation, project embedding representation, behavior embedding representation and project correlation embedding representation, the user's future behavior on the project is predicted.

[0133] The present invention provides a computer-readable storage medium, on which computer program instructions are stored. When the computer program instructions are executed by a processor, preset user embedding vectors, project embedding vectors and behavior embedding vectors are obtained as input features of a graph convolutional network; based on a message passing architecture of a graph convolutional neural network, graph structure information of heterogeneous interactions between users and projects and collaborative signals of multiple behaviors are captured through a user-project embedding propagation layer with behavior awareness, and the propagated user embedding representation, project embedding representation and behavior embedding representation are calculated; the same type of behavior interaction is performed on any two projects, the common features and associations between the any two projects are obtained, and the project correlation embedding representation is calculated; based on the user embedding representation, project embedding representation, behavior embedding representation and project correlation embedding representation, the user's future behavior on the project is predicted.

[0134] It should be understood that the specific features, operations and details described hereinabove about the method of the present invention may also be similarly applied to the device and system of the present invention, or, vice versa. In addition, each step of the method of the present invention described above may be performed by the corresponding parts or units of the device or system of the present invention.

[0135] It should be understood that each module / unit of the device of the present invention can be implemented in whole or in part by software, hardware, firmware or a combination thereof. Each module / unit can be embedded in the processor of the computer device in the form of hardware or firmware or independent of the processor, or can be stored in the memory of the computer device in the form of software for the processor to call to perform the operation of each module / unit. Each module / unit can be implemented as an independent component or module, or two or more modules / units can be implemented as a single component or module.

[0136] In one embodiment, a computer device is provided, which includes a memory and a processor, and the memory stores computer instructions executable by the processor, and the computer instructions instruct the processor to execute each step of the method of the embodiment of the present invention when executed by the processor. The computer device can be a server, a terminal, or any other electronic device with necessary computing and / or processing capabilities in a broad sense. In one embodiment, the computer device may include a processor, a memory, a network interface, a communication interface, etc. connected through a system bus. The processor of the computer device can be used to provide necessary computing, processing and / or control capabilities. The memory of the computer device may include a non-volatile storage medium and an internal memory. An operating system, a computer program, etc. may be stored in or on the non-volatile storage medium. The internal memory can provide an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface and the communication interface of the computer device can be used to connect and communicate with external devices through a network. The steps of the method of the present invention are executed by the processor.

[0137] The present invention may be implemented as a computer-readable storage medium having a computer program stored thereon, which causes the steps of the method of an embodiment of the present invention to be executed when executed by a processor. In one embodiment, the computer program is distributed on a plurality of computer devices or processors coupled to a network so that the computer program is stored, accessed, and executed in a distributed manner by one or more computer devices or processors. A single method step / operation, or two or more method steps / operations, may be performed by a single computer device or processor or by two or more computer devices or processors. One or more method steps / operations may be performed by one or more computer devices or processors, and one or more other method steps / operations may be performed by one or more other computer devices or processors. One or more computer devices or processors may perform a single method step / operation, or perform two or more method steps / operations.

[0138] It will be appreciated by those skilled in the art that the method steps of the present invention can be completed by instructing related hardware such as a computer device or a processor through a computer program, and the computer program can be stored in a non-temporary computer-readable storage medium, which causes the steps of the present invention to be executed when the computer program is executed. Depending on the circumstances, any reference to memory, storage, database or other media herein may include non-volatile and / or volatile memory. Examples of non-volatile memory include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (which significantly improves the accuracy and robustness of the recommendation system, making it suitable for various complex scenarios PROM), flash memory, magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid state disk, etc. Examples of volatile memory include random access memory (RAM), external cache memory, etc.

[0139] The various technical features described above can be combined arbitrarily. Although all possible combinations of these technical features are not described, any combination of these technical features should be considered to be covered by this specification as long as there is no contradiction in such combination.

[0140] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A graph convolutional network recommendation method integrating multiple behaviors, characterized in that: include: Obtain the preset user embedding vector, item embedding vector, and behavior embedding vector as input features of the graph convolutional network; Based on the message passing architecture of graph convolutional neural network, the user-item embedding propagation layer with behavior awareness captures the graph structure information of heterogeneous user-item interactions and the collaborative signals of multiple behaviors, and calculates the user embedding representation, item embedding representation and behavior embedding representation after propagation; Perform the same type of behavioral interactions on any two items and calculate the item correlation embedding representation; The user's future behavior on the project is predicted based on the user embedding representation, the project embedding representation, the behavior embedding representation and the project relevance embedding representation.

2. The graph convolutional network recommendation method integrating multiple behaviors as claimed in claim 1, characterized in that: The message passing architecture based on graph convolutional neural network captures the graph structure information of heterogeneous user-item interactions and the collaborative signals of multiple behaviors through a behavior-aware user-item embedding propagation layer, including: Based on the message passing architecture of graph convolutional neural network, a behavior-aware user-item embedding propagation layer is established; The user-item embedding propagation layer captures the graph structure information of user heterogeneous interactions and the collaborative signals of multiple behaviors.

3. The graph convolutional network recommendation method integrating multiple behaviors as described in claim 2, characterized in that: The user embedding representation includes: Rewrite the convolutional layer propagation formula of the graph convolutional neural network, embed and combine the relationship between adjacent project nodes to model the interaction behavior between users and projects; Different interactive behaviors of the user are assigned respective propagation weights, wherein the propagation weights integrate the intensity and quantity of the behaviors.

4. The graph convolutional network recommendation method integrating multiple behaviors as claimed in claim 1, characterized in that: The step of obtaining a preset user embedding vector, an item embedding vector, and a behavior embedding vector as input features of the graph convolutional network includes: The user embedding matrix U is composed of n user embedding vectors, the item embedding matrix V is composed of m item embedding vectors, and the behavior embedding matrix E is composed of t behavior embedding vectors, where the number of behavior types is determined by the type of interaction between users and items; The user embedding matrix U, the item embedding matrix V and the behavior embedding matrix E are used as input features of the graph convolutional network.

5. The graph convolutional network recommendation method integrating multiple behaviors as claimed in claim 1, characterized in that: The same type of behavioral interaction is performed on any two items, including: Performing the same type of behavioral interaction on any two items to obtain common features and associations between the two items; It captures the similarity or correlation of different behaviors in items and aggregates the features of items of the same type into the item embedding vector.

6. The graph convolutional network recommendation method integrating multiple behaviors as claimed in claim 1, characterized in that: The predicted user's future behavior on the project includes: Through propagation of a preset number of layers, multiple embedding representations of users, items, and behaviors are obtained and aggregated using preset weights; Calculate the prediction score of any user performing a certain action on any item. The prediction score not only captures the high-order collaborative signals of users and items under multiple actions, but also represents the semantics of the actions. The prediction score is optimized by minimizing a preset loss function.

7. A graph convolutional network recommendation device integrating multiple behaviors, characterized in that: include: An acquisition module is used to obtain preset user embedding vectors, item embedding vectors, and behavior embedding vectors as input features of the graph convolutional network; The first computing module is used for the message passing architecture based on graph convolutional neural network. Through the behavior-aware user-item embedding propagation layer, it captures the graph structure information of user-item heterogeneous interactions and the collaborative signals of multiple behaviors, and calculates the propagated user embedding representation, item embedding representation and behavior embedding representation; The second calculation module is used to perform the same type of behavioral interaction on any two items and calculate the item correlation embedding representation; A prediction module is used to predict the user's future behavior on the project based on the user embedding representation, the project embedding representation, the behavior embedding representation and the project relevance embedding representation.

8. The graph convolutional network recommendation device integrating multiple behaviors as claimed in claim 7, characterized in that: The acquisition module, the first calculation module, the second calculation module and the prediction module are controlled to execute the graph convolutional network recommendation method for integrating multiple behaviors as described in any one of claims 1-6.

9. An electronic device, characterized in that: include: Communication interface, processor, memory; Wherein, the memory is used to store program instructions, and when the program instructions are executed by the processor that is communicatively connected to the memory through the communication interface, the electronic device implements the graph convolutional network recommendation method that integrates multiple behaviors as described in any one of claims 1 to 6.

10. A computer-readable storage medium having program instructions stored thereon, characterized in that: When the program instructions are executed by a computer, the computer implements the graph convolutional network recommendation method for integrating multiple behaviors as described in any one of claims 1 to 6.

Images