A recommendation method based on multi-behavior information mining

Abstract

A recommendation method based on multi-behavior information mining, comprising: first, converting user and item features into low-dimensional dense vectors through embedding layer processing, providing basic feature input for capturing user behavior patterns. Secondly, using a combinatorial optimization graph convolutional network (COGCN) to capture user behavior patterns, and achieving multi-behavior fusion by imposing different degrees of constraints in the three stages before, during, and after the behavior. Then, using a distribution fitting multi-expert network (DFME) to coordinate the relationship between different tasks, and using contrastive learning to adaptively align the distribution of target behaviors and auxiliary behaviors to reduce feature distribution deviations. Finally, the outputs of DFME's behavior-specific experts and fitting experts are aggregated through a gating mechanism to form the final result of multi-behavior prediction and optimize prediction performance. The present invention has been comprehensively experimented on real-world datasets, verifying its superior performance in multi-behavior recommendation, and has great commercial value and application prospects.

Description

Technical Field

[0001] The present invention relates to intelligent recommendation technology, and in particular to a recommendation method based on multi-behavior information mining. Background Art

[0002] Recommender systems are a key technology in today's Internet services, providing high-quality personalized recommendations based on user preferences. Traditional collaborative filtering (CF) techniques learn user and item representations to make predictions and provide better recommendation results. Although these methods are effective to a certain extent, they only consider a single type of interaction between users and items, limiting their effectiveness in practical scenarios. In the real world, user behaviors are diverse. In addition to target behaviors (such as purchases), which are the main focus of businesses and platforms, users also participate in behaviors such as viewing, adding to cart, and favorites. This behavioral information contains different dimensions of user preferences and can serve as auxiliary knowledge to enhance the learning of target behaviors and provide better services to users.

[0003] To fully leverage auxiliary behaviors, multi-behavior recommendation methods have emerged. Existing multi-behavior recommendation methods can be divided into two steps: multi-behavior fusion and multi-behavior prediction. In the fusion step, multiple behaviors are combined to learn representations that capture user preferences. In the prediction step, these learned representations are applied to model prediction, with multi-task learning (MTL) demonstrating its effectiveness.

[0004] With the explosive development of deep learning, most methods for multi-behavior fusion have shifted from traditional matrix factorization to deep neural networks. Among them, graph neural networks (GNNs) have become a popular technique for multi-behavior recommendation due to their ability to effectively model high-order interactions between users and items. For example, MBGCN and CIGF simultaneously learn different behavioral information and fuse them through learnable parameters. Furthermore, some studies (e.g., MB-CGCN, CRGCN) leverage the cascading dependencies between real-world behaviors (e.g., view → shopping cart → purchase) to enhance model learning. BCIPM strategically models different behaviors, highlighting the importance of the target behavior. These methods essentially share the same goal: to capture richer user preferences by deeply exploring complex user behavior patterns formed by multiple interactions and the dependencies between behaviors. However, previous methods either simply aggregated behavior representations without constraints or defined strict sequential relationships for behaviors, indicating that user behavior patterns were not fully modeled, resulting in poor performance. In the multi-behavior prediction step, to address the limitation that a single label cannot fully represent the multiple preferences of users, multi-task learning (MTL) modules have been widely adopted, such as MMOE and PLE. These modules incorporate auxiliary behavior labels into joint optimization, allowing the model to leverage multi-behavior information to improve accuracy. To better adapt to multi-behavior prediction tasks, CIGF improves on the traditional MTL model by further decoupling the inputs of different tasks to reduce gradient conflicts. PKEF further builds on this by introducing a projection mechanism during aggregation to prevent the inclusion of harmful information.

[0005] Although the above multi-behavior recommendation methods have shown effectiveness in the multi-behavior fusion and prediction steps, the following challenges still exist:

[0006] Combinatorial Optimization of Multi-Behavior Fusion. Combinatorial optimization problems typically involve numerous states and choices, requiring the optimal solution to be found within bounded complexity. Multi-behavior recommendation is essentially a combinatorial optimization problem. For a given behavior category, each user has a finite number of possible behavior patterns. The optimal behavior pattern can be generated by exhaustively enumerating all possibilities, but this incurs significant space and time costs, known as "combinatorial explosion." Therefore, leveraging existing knowledge to constrain the solution space and achieve efficient multi-behavior recommendation is a challenge. Early methods focused solely on learning behavior-specific information without considering user behavior patterns. This was improved by methods that aggregated behaviors during the learning process, but they still lacked appropriate constraints on user behavior patterns. Some recent methods use a cascade paradigm to model user behavior sequences, resulting in overly strict constraints on user behavior patterns. In summary, there is an urgent need to propose a paradigm that establishes appropriate constraints to constrain the solution space.

[0007] Coordination of relationships between tasks. Multi-task learning (MTL) is a commonly used method in multi-behavior prediction, which models different behaviors as independent tasks. Since each task may affect the final prediction, it is crucial to correctly coordinate the relationship between tasks, which is currently limited by two factors: 1) Differences in feature space distribution during forward propagation: Most existing methods ignore this, resulting in bias in information aggregation. Although recent methods are dedicated to reducing feature distribution inconsistencies, such as projection-based aggregation mechanisms, they do not address the problem of representation space dynamics during training. 2) Differences in label space distribution during backpropagation: Existing methods learn information from different behaviors, but encounter conflicts during gradient updates due to differences in label distributions. Previous methods have attempted to use decoupled inputs to solve the gradient conflict problem, but the problem of gradient coupling during aggregation still exists. These factors may lead to negative transfer problems.

[0008] It should be noted that the information disclosed in the above background technology section is only used to understand the background of this application, and therefore may include information that does not constitute prior art known to ordinary technicians in this field. Summary of the Invention

[0009] The main purpose of the present invention is to overcome the defects existing in the above-mentioned background technology and provide a recommendation method based on multi-behavior information mining.

[0010] To achieve the above object, the present invention adopts the following technical solutions:

[0011] A recommendation method based on multi-behavior information mining includes the following steps:

[0012] S1. Embedding layer processing: The user and item features are converted into low-dimensional dense vector representations through the embedding layer, providing basic feature input for capturing user behavior patterns;

[0013] S2. Information Extraction via Combinatorial Optimized Graph Convolutional Networks: We use a Combinatorial Optimized Graph Convolutional Network (COGCN) to capture user behavior patterns. We implement multi-behavior fusion by imposing varying degrees of constraints at the pre-behavior, mid-behavior, and post-behavior stages. Pre-behavior constraints aggregate upstream behavior information to provide context for the current behavior. Mid-behavior constraints restrict the learning process of the current behavior, preventing the leakage of downstream behavior information to capture complex user behavior patterns. Post-behavior constraints independently output representations of each behavior for subsequent multi-task learning.

[0014] S3. Distribution Fitting Multi-Expert Network Task Coordination: The Distribution Fitting Multi-Expert Network (DFME) coordinates the relationships between different tasks, including using contrastive learning to adaptively align the distributions of target and auxiliary behaviors to reduce feature distribution bias. The Distribution Fitting Multi-Expert Network defines behavior-specific experts and fitting experts for each task. The behavior-specific experts are responsible for extracting features related to specific behaviors, while the fitting experts refine the representation space of each behavior through a graph convolutional network and fit the behavior feature distribution to reduce distribution bias between different behaviors and enhance the extraction and representation of behavior features.

[0015] S4. Multi-behavior prediction: The outputs of the behavior-specific experts and fitting experts of the distributed fitting multi-expert network (DFME) are aggregated through a gating mechanism to form the final result of multi-behavior prediction. The gating mechanism is used to balance the contributions of different experts and optimize the prediction performance, thereby combining the deep representation of user behavior patterns provided by COGCN with the task relationship and feature representation coordinated by DFME to jointly achieve multi-behavior prediction and generate recommendation results.

[0016] Furthermore, step S1 specifically includes:

[0017] Define user sets and item sets, as well as the behavior types and interaction matrix between users and items. The interaction matrix represents the interaction between users and items under different behavior types.

[0018] Retrieve low-dimensional dense embeddings of users and one-hot vectors of items from the embedding table to capture the feature representations of users and items under target and auxiliary behaviors.

[0019] Furthermore, in step S2, the combined optimization graph convolutional network includes the following mechanisms:

[0020] Define user behavior patterns: Define the high-frequency behavior chains between users and all items as user behavior patterns, which represent the user's personalized habits;

[0021] Define upstream and downstream behaviors: Based on user behavior habits, determine the order of behaviors and distinguish between upstream and downstream behaviors of the current behavior;

[0022] Imposing pre-behavior constraints: By aggregating upstream behavior information, providing context for the current behavior, constraining the interactions between behaviors, and ensuring that the input representation of the current behavior can reflect the influence of its upstream behavior;

[0023] Imposing behavioral constraints: When modeling current behavior, limit the inclusion of downstream behavior information to capture complex user behavior patterns and prevent information leakage;

[0024] Post-behavior constraints: Design decoupled outputs for each behavior, partition the solution space, and improve model performance through joint optimization.

[0025] Furthermore, in step S2, the implementation mechanism of the combined optimization graph convolutional network includes:

[0026] Utilize the graph convolutional network (GCN) paradigm to effectively simulate high-order connections between users and items to fuse multiple behavioral information;

[0027] Adjust the adjacency matrix of different behaviors to adapt to the requirements of graph convolution and achieve effective information transfer between behaviors;

[0028] Capture the interaction information between users and items in each behavior through graph convolution to achieve behavior constraints;

[0029] In each layer of the graph convolutional network, information about the current behavior and its upstream behavior is combined to perform message propagation and aggregation to implement behavior constraints;

[0030] Define a hierarchical information transfer mechanism between behaviors, combining the high-level representation of upstream behaviors with the initial representation of the current behavior as input to implement pre-behavior constraints;

[0031] The representation of each behavior is output independently to meet the needs of multi-task learning, and rich relational representations are obtained by stacking the outputs of different layers, following the post-behavior constraints.

[0032] Furthermore, in step S3, the distribution fitting multi-expert network DFME includes the following contrastive learning mechanism:

[0033] Utilize contrastive learning to adaptively learn the distribution similarity between target and auxiliary behaviors to mitigate the distribution bias between different data sources;

[0034] Applying the temperature hyperparameter of the softmax function and the similarity calculation function, comparative learning is performed on the user and item sides to generate behavior-specific experts;

[0035] The behavior representation is decoupled using the Hadamard product operation to prevent gradient conflicts caused by coupled inputs, thereby generating multiple behavior-specific experts.

[0036] Furthermore, in step S3, the distribution fitting multi-expert network DFME includes the following behavior fitting expert generation mechanism:

[0037] Define a dedicated fitting expert for each task to solve the problem of feature distribution deviation, and its output is used for subsequent aggregation process;

[0038] By controlling the scaling factor of the behavior representation, fine-tuning of the representation space is achieved;

[0039] Use graph convolution operations to capture the effective information contained in the behavior interaction matrix to enhance the extraction and representation of behavior features;

[0040] Use graph convolutional networks to process behavior representations and achieve effective information transfer between behaviors through layer-specific parameters;

[0041] The final expert representation is generated through the Hadamard product operation, similar to the behavior-specific expert, to achieve behavior fitting.

[0042] Furthermore, in step S3, the distribution fitting multi-expert network DFME includes the following task aggregation mechanism:

[0043] Define the gating of the task, use the feature transformation matrix and bias matrix, and the attention vector as selectors, and calculate the weighted sum of all experts;

[0044] Utilize behavior-aware graph convolution operations and representation scaling mechanisms to capture the effective components of other behaviors to alleviate the negative transfer caused by feature distribution differences in gated aggregation;

[0045] By refining the representation and the representation of the target behavior, target aggregation is performed while ensuring that the parameters of the target task are not affected by the gradient update of the auxiliary task and preventing interference from the auxiliary behavior;

[0046] Apply a stop gradient operation to ensure that the auxiliary behavior does not negatively impact the prediction of the target behavior;

[0047] The final prediction for task k is formed by weighted summing the outputs of all actions;

[0048] The BPR loss optimization model is used to enhance the recommendation accuracy of the model by comparing the predicted values ​​of positive and negative samples.

[0049] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the recommendation method based on multi-behavior information mining.

[0050] A computer program product includes a computer program, wherein when the computer program is executed by a processor, the recommendation method based on multi-behavior information mining is implemented.

[0051] The present invention has the following beneficial effects:

[0052] This paper proposes a recommendation method based on multi-behavior information mining. It designs a multi-behavior recommendation framework (COPF) based on a combinatorial optimization perspective. This framework comprises a combinatorial optimization graph convolutional network (COGCN) and a distribution fitting multi-expert network (DFME). To address the combinatorial optimization problem in the fusion step, COGCN imposes varying degrees of constraints on user behavior patterns at various stages (before, during, and after the behavior) to constrain the combinatorial optimization solution space, thereby achieving efficient multi-behavior fusion. To coordinate the relationships between tasks, DFME improves the forward and backpropagation processes in the multi-behavior prediction phase from both feature and label perspectives. At the feature level, DFME treats different behaviors as independent tasks and uses contrastive learning to adaptively align the distributions of target and auxiliary behaviors. Considering that behavior aggregation may be affected by differences in behavior feature distributions, DFME incorporates a dedicated fitting expert before aggregation to refine the representation space of each behavior, thereby reducing distribution bias while maintaining spatial generalization. At the label level, DFME further decouples the gradients between the target and auxiliary behaviors during aggregation to prevent the influence of other tasks on the target task gradient update. The above design makes it possible to effectively use auxiliary tasks to adjust the data distribution of the model fitting to make it more consistent with the test distribution of the target task, thereby alleviating the negative transfer problem caused by the incoordination of task relationships.

[0053] The present invention innovatively examines the multi-behavior fusion process from the perspective of combinatorial optimization, and effectively promotes multi-behavior fusion by imposing different degrees of constraints at various stages of the user behavior pattern, thereby improving the recommendation system's ability to capture user preferences. The method of the present invention not only overcomes the limitations of existing multi-behavior recommendation methods in failing to fully mine user behavior pattern information during the fusion process, but also solves the negative transfer problem in multi-task learning through the distributed fitting multi-expert network (DFME), and optimizes the multi-behavior prediction process. The recommendation method of the present invention has demonstrated superior performance on real-world datasets, proving its rationality and effectiveness. Compared with traditional recommendation systems, the present invention can comprehensively analyze a variety of user behaviors, reasonably capture the correlation between behaviors and user behavior patterns through deep learning methods, and provide more accurate recommendations. In addition, the multi-task learning scheme of the present invention can coordinate the relationship between tasks, alleviate the negative transfer problem, achieve more accurate predictions, and has efficient computing performance, which is of great significance to practical application scenarios.

[0054] Other beneficial effects of the embodiments of the present invention will be further described below. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a basic flow chart of the recommendation method based on multi-behavior information mining of the present invention.

[0056] Figure 24 is a flowchart of a recommended method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] The following is a detailed description of the embodiments of the present invention. It should be emphasized that the following description is only exemplary and is not intended to limit the scope of the present invention and its application.

[0058] A recommendation method based on multi-behavior information mining aims to solve the problem of insufficient user behavior pattern mining in existing recommendation systems. The key points of the method include: First, user and item features are converted into low-dimensional dense vectors through embedding layer processing, providing basic feature input for capturing user behavior patterns. Secondly, the combined optimization graph convolutional network COGCN is used to capture user behavior patterns, and multi-behavior fusion is achieved by imposing different degrees of constraints in the three stages before, during, and after the behavior. Then, the distribution fitting multi-expert network DFME is used to coordinate the relationship between different tasks, and contrastive learning is used to adaptively align the distribution of target behavior and auxiliary behavior to reduce feature distribution bias. Finally, the outputs of DFME's behavior-specific experts and fitting experts are aggregated through a gating mechanism to form the final result of multi-behavior prediction and optimize prediction performance.

[0059] See Figure 1 and Figure 2 , an embodiment of the present invention provides a recommendation method based on multi-behavior information mining, comprising the following steps:

[0060] Step S1. Embedding layer processing: The user and item features are converted into low-dimensional dense vector representations through the embedding layer, providing basic feature input for subsequent user behavior pattern capture;

[0061] In a preferred embodiment, step S1 specifically includes: defining a set of users and a set of items, as well as the behavior types and interaction matrix between users and items, where the interaction matrix represents the interaction of users with items under different behavior types; retrieving the low-dimensional dense embedding of users and the one-hot vector of items from the embedding table to capture the feature representation of users and items under target behavior and auxiliary behavior.

[0062] Step S2. Combinatorial Optimization Graph Convolutional Network Information Extraction: Utilizing the Combinatorial Optimization Graph Convolutional Network (COGCN), user behavior patterns are captured. By imposing different degrees of constraints at the pre-behavior, mid-behavior, and post-behavior stages, efficient multi-behavior fusion is achieved. Pre-behavior constraints aggregate upstream behavior information to provide context for the current behavior; mid-behavior constraints restrict the learning process of the current behavior to prevent the leakage of downstream behavior information, thereby capturing more complex user behavior patterns; and post-behavior constraints independently output representations of each behavior, preparing for subsequent multi-task learning.

[0063] In a preferred embodiment, in step S2, the combinatorial optimization graph convolutional network includes the following mechanisms: defining user behavior patterns: defining high-frequency behavior chains between users and all items as user behavior patterns, representing the user's personalized habits; defining upstream and downstream behaviors: determining the sequential relationship between behaviors based on user behavior habits, and distinguishing the upstream and downstream behaviors of the current behavior; imposing pre-behavior constraints: providing context for the current behavior by aggregating upstream behavior information, constraining the interaction between behaviors, and ensuring that the input representation of the current behavior can reflect the influence of its upstream behavior; imposing mid-behavior constraints: when modeling the current behavior, limiting the inclusion of downstream behavior information to capture complex user behavior patterns and prevent information leakage; imposing post-behavior constraints: designing decoupled outputs for each behavior, dividing the solution space, and improving model performance through joint optimization.

[0064] In a preferred embodiment, in step S2, the implementation mechanism of the combined optimized graph convolutional network includes: using the paradigm based on the graph convolutional network GCN to effectively simulate high-order connections between users and items to fuse multi-behavior information; adjusting the adjacency matrix of different behaviors to adapt to the requirements of graph convolution to achieve effective transmission of information between behaviors; capturing the interaction information of users and items under each behavior through graph convolution to achieve constraints in behavior; in each layer of the graph convolutional network, combining the information of the current behavior and its upstream behavior, performing message propagation and aggregation to achieve constraints in behavior; defining a hierarchical information transmission mechanism between behaviors, combining the high-level representation of the upstream behavior with the initial representation of the current behavior as input to achieve pre-behavior constraints; independently outputting the representation of each behavior to meet the needs of multi-task learning, obtaining rich relationship representations by superimposing the outputs of different layers, and following post-behavior constraints.

[0065] Step S3. Distribution Fitting Multi-Expert Network Task Coordination: The distribution fitting multi-expert network (DFME) is used to coordinate the relationships between different tasks, including using contrastive learning to adaptively align the distributions of target and auxiliary behaviors to reduce feature distribution bias. The distribution fitting multi-expert network defines behavior-specific experts and fitting experts for each task. The behavior-specific experts are responsible for extracting features related to specific behaviors, while the fitting experts refine the representation space of each behavior through a graph convolutional network and fit the behavior feature distribution to reduce the distribution bias between different behaviors and enhance the extraction and representation of behavior features, thereby meeting the needs of multi-task learning.

[0066] In a preferred embodiment, the distribution fitting multi-expert network DFME includes the following contrastive learning mechanism: using contrastive learning to adaptively learn the distribution similarity between target behavior and auxiliary behavior to alleviate the distribution deviation between different data sources; applying the temperature hyperparameters and similarity calculation function of the softmax function to perform contrastive learning on the user and item sides to generate behavior-specific experts; using the Hadamard product operation to decouple the behavior representation to prevent gradient conflicts caused by coupled inputs, thereby generating multiple behavior-specific experts.

[0067] In a preferred embodiment, the distribution fitting multi-expert network DFME includes the following behavior fitting expert generation mechanism: defining a special fitting expert for each task to solve the problem of feature distribution deviation, and its output is used for the subsequent aggregation process; by controlling the coefficient of behavior representation scaling, fine-tuning of the representation space is achieved; using graph convolution operations to capture the effective information contained in the behavior interaction matrix to enhance the extraction and representation of behavior features; using graph convolution networks to process the behavior representation, and realizing effective information transfer between behaviors through layer-specific parameters; generating the final expert representation through Hadamard product operations, similar to behavior-specific experts, to achieve behavior fitting.

[0068] In a preferred embodiment, the distribution fitting multi-expert network DFME includes the following task aggregation mechanism: defining the gating of the task, using the feature transformation matrix and bias matrix, and the attention vector as a selector, and calculating the weighted sum of all experts; utilizing the behavior-aware graph convolution operation and representation scaling mechanism to capture the effective components of other behaviors to alleviate the negative transfer caused by the difference in feature distribution in the gated aggregation; performing target aggregation by refining the representation and the representation of the target behavior, while ensuring that the parameters of the target task are not affected by the gradient update of the auxiliary task to prevent interference from the auxiliary behavior; applying the stop gradient operation to ensure that the auxiliary behavior does not have a negative impact on the prediction of the target behavior; forming the final prediction of the task by weighted summing the outputs of all behaviors; using the BPR loss optimization model to enhance the recommendation accuracy of the model by comparing the predicted values ​​of positive and negative samples.

[0069] Step S4. Multi-behavior prediction: The outputs of the behavior-specific experts and fitting experts of the distributed fitting multi-expert network (DFME) are aggregated through a gating mechanism to form the final result of multi-behavior prediction. The gating mechanism is used to balance the contributions of different experts and optimize the prediction performance, thereby combining the deep representation of user behavior patterns provided by COGCN and the task relationship and feature representation coordinated by DFME to jointly achieve multi-behavior prediction and generate more accurate recommendation results.

[0070] The present invention applies a unique combinatorial optimization perspective to multi-behavior recommendation systems to solve the problem that existing technologies fail to fully mine user behavior pattern information. Through the combinatorial optimization graph convolutional network (COGCN), the present invention imposes different degrees of constraints at various stages of the user behavior pattern, effectively promoting multi-behavior fusion and enhancing the model's ability to capture complex user behavior patterns; the introduction of the distribution fitting multi-expert network (DFME) addresses the negative transfer problem in multi-task learning, coordinates the relationship between different tasks by improving the forward and backward propagation processes, and thus optimizes the multi-behavior prediction process. Compared with traditional recommendation systems, the present invention can comprehensively analyze multiple user behaviors, accurately capture the correlation between behaviors and user behavior patterns through deep learning methods, provide more accurate recommendations, and at the same time alleviate the negative transfer problem in multi-task learning, achieving more accurate predictions.

[0071] The following further describes an algorithm example and experimental verification of a specific embodiment of the present invention.

[0072] To address the two challenges mentioned above, this paper proposes a recommendation method based on multi-behavior information mining and designs a multi-behavior recommendation framework (COPF) based on a combinatorial optimization perspective. This framework comprises a combinatorial optimization graph convolutional network (COGCN) and a distribution fitting multi-expert network (DFME). To address the combinatorial optimization problem in the fusion step, COGCN imposes varying degrees of constraints on user behavior patterns at various stages (before, during, and after the behavior) to constrain the combinatorial optimization solution space, thereby achieving efficient multi-behavior fusion. To coordinate the relationships between tasks, DFME improves the forward and backward propagation processes in the multi-behavior prediction phase from both feature and label perspectives. At the feature level, DFME treats different behaviors as independent tasks and uses contrastive learning to adaptively align the distributions of target and auxiliary behaviors. Considering that behavior aggregation may be affected by differences in behavior feature distributions, DFME incorporates a dedicated fitting expert before aggregation to refine the representation space of each behavior, thereby reducing distribution bias while maintaining spatial generalization. At the label level, DFME further decouples the gradients between the target and auxiliary behaviors during aggregation to prevent the influence of other tasks on the target task gradient updates. The above design makes it possible to effectively use auxiliary tasks to adjust the data distribution of the model fitting to make it more consistent with the test distribution of the target task, thereby alleviating the negative transfer problem caused by the incoordination of task relationships.

[0073] This paper designs a "Combinatorial Optimization Perspective-based Framework" (COPF) for multi-behavior recommendation, which consists of two parts: (1) Combinatorial Optimization Graph Convolutional Network (COGCN); (2) Distribution Fitting Multi-Expert Network (DFME). Figure 1 This is the basic process of the proposed recommendation method. Figure 2This is a flow chart of an embodiment of the present invention. Specific implementation examples are as follows:

[0074] 1) First, define the problem.

[0075] In the framework of the present invention, U and V are used to represent users and items respectively. The user set and item set are represented as U = {u1, u2, ..., u nU} and V={v1,v2,…,v nV}, where n U and n V Represent the total number of users and items respectively. The behavior type k∈{1,2,…,K} maintains a consistent order between behaviors (i.e., 1 and K correspond to the most upstream and most downstream behaviors, respectively). The user-item interaction matrix for k behavior types can be expressed as B={B1,B2,…,B K}. Among them B k ∈{0,1} indicates whether user u interacts with item v under behavior k. For multi-behavior recommendation, there is a target behavior that needs to be optimized (e.g., purchase), and other behaviors are considered auxiliary behaviors to assist in predicting the target behavior. The target behavior is the most downstream behavior (behavior K). Find the user's low-dimensional dense embedding e from the embedding table u and the item's e v , using their one-hot vectors.

[0076] 2) Combinatorial Optimization Graph Convolutional Network

[0077] Recent multi-behavior methods aim to deeply explore user behavior patterns to enhance model performance. However, the constraints they impose on user behavior patterns during the fusion step are either too strict or too loose, making it difficult for the model to accurately capture these patterns. To address this issue, we examine the multi-behavior fusion process from a combinatorial optimization perspective and propose a combinatorial optimization graph convolutional network (COGCN). It imposes varying degrees of constraints at different stages of user behavior in order to learn optimal behavior information. To better describe the constraints at different behavior stages, we first introduce the following definitions:

[0078] Definition 1 (User Behavior Pattern). A user behavior pattern is defined as a high-frequency behavior chain between a user and all items on the platform, representing the user's personalized habits. Formally, for any user u∈U, his / her behavior pattern is: u→b1→…→b k →v, where b k is the kth behavior, and K is the total number of behavior types. For example, for a user who buys an attractive item directly after browsing it, his / her behavior pattern is: u→view→buy→v.

[0079] Definition 2 (Upstream and Downstream Behaviors). The order of behaviors is typically derived from the prevailing behavior habits of most users in the real world. For example, consider the combination view → cart → buy. An upstream behavior is the behavior that precedes the current behavior. Similarly, a downstream behavior is the behavior that follows the current behavior. In this example, view is the upstream behavior of cart, and buy is the downstream behavior of cart.

[0080] As can be seen, as the number of behavior types increases, the total number of possible user behavior patterns explodes. To limit the solution space for combinatorial optimization in a multi-behavior recommendation system, this paper imposes constraints on user behavior patterns at three stages: before, during, and after the behavior. Details are as follows.

[0081] Definition 3 (Before-action Constraints). Many user-item interactions follow a predetermined sequence of actions. Therefore, the upstream behavior information of the current action is crucial. Formally, for any action, its input is expressed as:

[0082]

[0083] f(·) is an aggregation function (e.g., summation) and g(·) is a transformation function (e.g., matrix multiplication). In this way, the interactions between behaviors are constrained.

[0084] Definition 4 (Constraints in Behavior). Constraints in behavior essentially involve modeling the current behavior (i.e., the transformation function g(·)). In order to capture more complex or implicit user behavior patterns (e.g., users alternate between viewing and adding to the shopping cart before purchasing), GCNs with heterogeneous relationships are also utilized. At the same time, to prevent poor model generalization performance and overfitting problems due to information leakage, the learning process of the current behavior node is defined to not contain semantic information of downstream behaviors. Specifically, for the current behavior, we have:

[0085]

[0086] Definition 5 (Post-Behavior Constraints): For each behavior, decoupled outputs are designed to partition the solution space. Through joint optimization, the model can achieve better performance. This is consistent with the DFME concept proposed in this paper.

[0087] In the previous section, we outlined the specific constraints for different behavior stages. Since graph convolutional networks (GCNs) can effectively exploit high-order connections between users and items, we use a GCN-based paradigm to model multi-behavior information fusion from a combinatorial optimization perspective. Given the adjacency matrices of different behaviors, we modify them to meet the requirements of graph convolution:

[0088]

[0089] Among them A k is the adjacency matrix of row k. For the same purpose, we obtain the embedding matrix of users and items

[0090]

[0091] Then, graph convolution is used to capture the interaction information of behaviors. For behavior k, we have:

[0092]

[0093] Where D is a diagonal matrix and I represents the identity matrix. We use the adjacency matrix of the current behavior and its upstream behavior to propagate and aggregate messages at each level, thus achieving intra-behavior constraints. Furthermore, to achieve pre-behavior constraints, we define the hierarchical information transfer between behaviors as follows:

[0094]

[0095] Where L represents the total number of GCN layers. Here, the final layer representation of each upstream behavior is combined with the initial representation as the input for the current behavior. Following the post-behavior constraints, the representation of each behavior is independently output for subsequent multi-task learning. Specifically, the outputs of different layers are directly added to obtain relationships of different orders.

[0096]

[0097] 3) Distributed Fitting Multi-Expert Network

[0098] By using COGCN in the multi-behavior fusion step, the representation of user u and item v under each behavior k has been obtained. The next task is to design an appropriate structure for multi-behavior prediction. Many methods have utilized multi-task learning (MTL) modules to make full use of multi-behavior information to assist in predicting target behaviors, which has shown their effectiveness. However, these MTL methods have insufficient exploration in structural design and fail to take into account the potential negative transfer effects that may be caused by differences in feature and label distribution during the learning process. In order to address the shortcomings of the existing MTL module, the present invention proposes a distributed fitting multi-expert network (DFME), which controls the interaction between behaviors through features and labels, thereby coordinating the relationship between tasks. The specific details are as follows.

[0099] 1. Since contrastive learning can alleviate the distribution bias between different data sources, it is used to adaptively learn the distribution similarity between the target behavior and auxiliary behaviors, and then regenerate the expert. Taking auxiliary behavior k as an example, we have:

[0100]

[0101] Where τ represents the temperature hyperparameter of the softmax function, is a function that calculates the similarity between two vectors (e.g., inner product). The item side follows the same contrastive learning process. The final loss is:

[0102]

[0103] Then, following the previous approach, we use disentangled behavior representations to generate behavior-specific experts, thus preventing gradient conflicts caused by coupled inputs:

[0104]

[0105] in is the Hadamard product operation. Since the decoupled behavior representation is used to generate experts, a total of k behavior-specific experts can be obtained.

[0106] 2. Generation of Behavior Fitting Experts. To address the challenge of mitigating feature distribution bias, a dedicated fitting expert is defined for each task, whose output is used in the subsequent aggregation process. Specifically, for the current behavior k and any other behavior k′, it can be expressed as:

[0107]

[0108] α and β are coefficients that control the scaling of the behavior representation, and Agg is the graph convolution operation. In particular, the values ​​of α and β should be small to achieve a fine-tuning effect on the representation space. The interaction matrix between the graph convolution network and the behavior k′ is used to capture the representation The effective information contained in the graph convolution operator is defined as follows:

[0109]

[0110] A k ′ is the adjacency matrix of row k′, W l are layer-specific parameters shared with the GCN layer. The Hadamard product operation is used to generate the final expert, which is similar to the behavior-specific expert.

[0111]

[0112] 3. Expert Aggregation. To mitigate the negative transfer problem caused by differences in feature and label distributions, we improved the task aggregation mechanism, including forward and backward propagation. As we can see, the behavior-aware graph convolution operation and representation scaling mechanism help capture the effective components of other behaviors, which can be further used to mitigate the negative transfer caused by differences in feature distribution in gated aggregation. Specifically, the gate for task k is defined as:

[0113]

[0114] Wg and bg are the feature conversion matrix and bias matrix, g k is the attention vector, used as a selector to calculate the weighted sum of all experts. Then take the refined representation of other behaviors

[0115] e j，k (j∈{1,2,...,K}∩j≠k)

[0116] and e k As the kth gate of the target aggregation. In order to eliminate the negative impact caused by the difference in label distribution between the auxiliary behavior and the target behavior, it is ensured that the parameters of the target task will not be updated by the gradient update of the auxiliary task, thereby preventing the interference of the auxiliary behavior. Formally, there is:

[0117]

[0118] where g k (j) represents the vector g k The jth element of . sg(·) is the stopping gradient operation. The final prediction of task k is

[0119]

[0120] Use BPR Loss to optimize the model:

[0121]

[0122] experiment

[0123] In order to verify the effect of the present invention, three datasets extracted from real scenarios (Beibei, Taobao and Tmall) are used to verify the effectiveness of the scheme proposed in this invention. The method proposed in this invention will be compared with many baseline models in recent years, which can be divided into three categories: (1) single behavior recommendation methods: MF-BPR, NeuMF and LightGCN, (2) multi-behavior recommendation methods without MTL: RGCN, GNMR, NMTR, MBGCN, S-MBRec, KMCLR and MB-CGCN, (3) multi-behavior recommendation methods with MTL: CML, CRGCN, CIGF, PKEF and BCIPM. In order to evaluate the performance of the proposed method and the baseline method in top-k recommendation, two indicators are used: Hit Ratio (HR@K) and Normalized Discounted Cumulative Gain (NDCG@K). And set K=10 in all experiments. The experimental results are shown in Table 1:

[0124] Table 1

[0125]

[0126] The proposed COPF model achieved state-of-the-art performance on all three datasets. Specifically, on the Beibei dataset, COPF improved the best baseline by 49.91% in HR and 55.15% in NDCG; on the Taobao dataset, COPF achieved improvements of 12.06% and 6.75%, respectively; and on the Tmall dataset, COPF achieved improvements of 24.12% and 30.50%, respectively. Given the diverse user behavior patterns across datasets, COPF's superior performance across all datasets further demonstrates its applicability and effectiveness in multi-behavior recommendation.

[0127] Among single-behavior methods, LightGCN outperforms MF-BPR and NeuMF, while among multi-behavior methods, MBGCN also outperforms NMTR. This demonstrates the advantage of GCNs in capturing high-order interaction information. In addition, most multi-behavior recommendation methods, such as MBGCN, outperform single-behavior methods on all three datasets, highlighting the advantages of leveraging multi-behavior information for learning. Finally, the superior performance of CML and KMCLR also demonstrates the effectiveness of contrastive learning.

[0128] Although the models differ in network structure, in general, multi-behavior methods with MTL perform better than those without MTL. For example, PKEF outperforms all multi-behavior methods without MTL on all datasets. It is worth noting that KMCLR and MB-CGCN perform better among multi-behavior methods without MTL. The possible reason is that KMCLR enhances the original multi-behavior information by introducing external knowledge graph information; at the same time, MB-CGCN reduces the solution space of the multi-behavior fusion problem through cascade constraints. Although this constraint is too strict, it still achieves relatively good results in the biased space.

[0129] GNMR and MBGCN outperform RGCN by considering the contribution of each behavior in behavior fusion. Compared with them, CIGF uses multi-task learning in the prediction process, which further improves the performance. However, they still lack appropriate constraints or impose overly loose constraints on user behavior patterns. Recent methods, such as CRGCN, MBCGCN and PKEF, use the cascade paradigm to constrain the learning of user behavior patterns; BCIPM further relaxes the constraints within the cascade paradigm and emphasizes the importance of target behavior. As can be seen, PKEF performs second only to the model of the present invention on the Beibei and Taobao datasets, while BCIPM does the same on the Tmall dataset. This shows the necessity of considering user behavior pattern constraints from a combinatorial optimization perspective.

[0130] Table 2

[0131]

[0132] The proposed DFME can serve as a universal module applicable to most existing multi-behavioral methods. To validate this assertion, we selected several representative multi-behavioral methods, such as CRGCN, CIGF, and PKEF, as well as multi-behavioral methods without MTL, such as LightGCN_M (LightGCN enhanced with a multi-behavioral graph network) and MB-CGCN. We then replaced their prediction modules with DFME and compared their performance with the original models. The results are shown in Table 2. As can be seen, the proposed DFME improves the performance of all the original models. The original LightGCN_M and MB-CGCN, due to their lack of an MTL module, benefit significantly from DFME. Compared to CIGF and PKEF, CRGCN shows more significant improvements, likely due to its simpler structure, better compatibility, and its relatively basic original MTL module. In contrast, CIGF and PKEF already have relatively comprehensive MTL modules, resulting in less significant improvements. Overall, the results fully demonstrate the broad compatibility and versatility of the proposed DFME with different multi-behavioral models, and can significantly improve the performance of the original models.

[0133] In order to evaluate the effectiveness of the submodules in the COPF framework proposed in this paper, ablation experiments were conducted on COGCN and DFME respectively. For COGCN, the constraints at each stage are mainly considered. Specifically, two strict constraints are first defined: (1) Strict pre-behavior constraint: the pre-behavior constraint is changed to "only receive information from the most recent upstream behavior"; (2) Strict mid-behavior constraint: the mid-behavior constraint is changed to "only learn the current behavior signal". Then the following variants are designed: (1) w / o COGCN: COGCN is replaced with multiple LightGCNs. (2) COPF-P: remove the pre-behavior constraint. (3) COPF-A: remove the mid-behavior constraint. (4) COPF-D: remove the post-behavior constraint (and no DFME). (5) COPF-F: remove both pre-behavior and mid-behavior constraints. (6) COPF-C: use a strict cascade paradigm. (7) COPF-B: use a strict pre-behavior constraint instead. (8) COPF-H: use a strict mid-behavior constraint instead. For DFME, there are: (1) w / o DFME: replacing DFME with a bilinear module; (2) w / o con.: removing contrastive learning. (3) w / o for.: removing the improvement of forward propagation during aggregation. (4) w / o back.: removing the improvement of backward propagation during aggregation. (5) w / o fit.: removing the improvement of bidirectional propagation during aggregation. The results in Table 3 lead to the following conclusions:

[0134] Table 3

[0135]

[0136] Comparing COPF with other variants that modify COGCN constraints (for fairness, COPF-D is compared with COPF without DFME), we observe that removing or changing constraints at any stage leads to varying degrees of performance degradation. Furthermore, COPF without DFME performs worse than other variants on all three datasets, demonstrating the effectiveness of approaching multi-behavior fusion from a combinatorial optimization perspective and validating the rationality of the constraints established by COGCN at each stage. The performance of each variant with DFME is also affected to varying degrees, with the variant without DFME performing the worst. This demonstrates the rationality and effectiveness of the DFME proposed in this paper.

[0137] Compared with traditional recommendation technology, the innovation and advantages of the present invention are as follows:

[0138] 1. This invention is the first to introduce combinatorial optimization into the multi-behavior recommendation problem and proposes to examine the multi-behavior fusion problem from the perspective of combinatorial optimization.

[0139] 2. This paper divides the multi-behavior recommendation method into two stages: multi-behavior fusion and multi-behavior prediction. It also proposes a multi-behavior recommendation framework (COPF) based on a combinatorial optimization perspective. This framework includes a combinatorial optimization graph convolutional network (COGCN) and a distribution fitting multi-expert network (DFME).

[0140] 3. Compared with the limitations of existing methods, this paper takes advantage of the benefits of behavioral constraints for multi-behavior recommendation and proposes a module called Combinatorial Optimization Graph Convolutional Network (COGCN) as a solution. By imposing different degrees of constraints on user behavior patterns at each stage, it effectively promotes the multi-behavior fusion process.

[0141] 4. This paper overcomes the limitations of existing multi-task methods in multi-behavior recommendation from a structural perspective (i.e., feature and label perspective), and proposes a distributed fitting multi-expert network (DFME), which coordinates the relationship between tasks by improving the forward and backward propagation processes, thereby alleviating the negative transfer problem in multi-task learning and optimizing the multi-behavior prediction process.

[0142] 5. The solution proposed in this invention is simple and easy to integrate, can achieve end-to-end learning, and is suitable for implementation in commercial scenarios.

[0143] 6. The solution proposed in the present invention can be applied in all stages of a real recommendation platform (such as rough ranking, fine ranking, etc.) and has strong usability.

[0144] 7. Comprehensive experiments are conducted on three real-world datasets, demonstrating the superior performance of the proposed COPF in multi-behavior recommendation. Further experimental results also verify the rationality and effectiveness of the COGCN and DFME modules.

[0145] An embodiment of the present invention further provides a storage medium for storing a computer program, which at least performs the above method when executed.

[0146] An embodiment of the present invention further provides a control device, comprising a processor and a storage medium for storing a computer program; wherein the processor is configured to execute at least the method described above when executing the computer program.

[0147] An embodiment of the present invention further provides a processor, which executes a computer program and at least performs the method described above.

[0148] The storage medium can be implemented by any type of non-volatile storage device, or a combination thereof. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a magnetic random access memory (FRAM), a flash memory (Flash Memory), a magnetic surface memory, an optical disc, or a read-only optical disc (CD-ROM); the magnetic surface memory can be a magnetic disk memory or a magnetic tape memory. The storage medium described in the embodiments of the present invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0149] In the several embodiments provided by the present invention, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0150] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0151] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0152] Those skilled in the art will understand that all or part of the steps of the above-mentioned method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above-mentioned method embodiment; and the aforementioned storage medium includes: mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, etc. Various media that can store program codes.

[0153] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

[0154] The methods disclosed in the several method embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0155] The features disclosed in several product embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new product embodiments.

[0156] The features disclosed in several method or device embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments or device embodiments.

[0157] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. Those skilled in the art will recognize that, without departing from the scope of the present invention, several equivalent substitutions or obvious variations can be made, and the performance or use of the same should be considered to fall within the scope of protection of the present invention.

Claims

1. A recommendation method based on multi-behavior information mining, characterized in that: The following steps are involved: S1. Embedding layer processing: The user and item features are converted into low-dimensional dense vector representations through the embedding layer, providing basic feature input for capturing user behavior patterns; S2. Information Extraction via Combinatorial Optimized Graph Convolutional Networks: We use a Combinatorial Optimized Graph Convolutional Network (COGCN) to capture user behavior patterns. We implement multi-behavior fusion by imposing varying degrees of constraints at the pre-behavior, mid-behavior, and post-behavior stages. Pre-behavior constraints aggregate upstream behavior information to provide context for the current behavior. Mid-behavior constraints restrict the learning process of the current behavior, preventing the leakage of downstream behavior information to capture complex user behavior patterns. Post-behavior constraints independently output representations of each behavior for subsequent multi-task learning. S3. Distribution Fitting Multi-Expert Network Task Coordination: The Distribution Fitting Multi-Expert Network (DFME) coordinates the relationships between different tasks, including using contrastive learning to adaptively align the distributions of target and auxiliary behaviors to reduce feature distribution bias. The Distribution Fitting Multi-Expert Network defines behavior-specific experts and fitting experts for each task. The behavior-specific experts are responsible for extracting features related to specific behaviors, while the fitting experts refine the representation space of each behavior through a graph convolutional network and fit the behavior feature distribution to reduce distribution bias between different behaviors and enhance the extraction and representation of behavior features. S4. Multi-behavior prediction: The outputs of the behavior-specific experts and fitting experts of the distributed fitting multi-expert network (DFME) are aggregated through a gating mechanism to form the final result of multi-behavior prediction. The gating mechanism is used to balance the contributions of different experts and optimize the prediction performance, thereby combining the deep representation of user behavior patterns provided by COGCN with the task relationship and feature representation coordinated by DFME to jointly achieve multi-behavior prediction and generate recommendation results.

2. The recommendation method based on multi-behavior information mining according to claim 1, characterized in that: Step S1 specifically includes: Define user sets and item sets, as well as the behavior types and interaction matrix between users and items. The interaction matrix represents the interaction between users and items under different behavior types. Retrieve low-dimensional dense embeddings of users and one-hot vectors of items from the embedding table to capture the feature representations of users and items under target and auxiliary behaviors.

3. The recommendation method based on multi-behavior information mining according to claim 1 or 2, characterized in that: In step S2, the combined optimization graph convolutional network includes the following mechanisms: Define user behavior patterns: Define the high-frequency behavior chains between users and all items as user behavior patterns, which represent the user's personalized habits; Define upstream and downstream behaviors: Based on user behavior habits, determine the order of behaviors and distinguish between upstream and downstream behaviors of the current behavior; Imposing pre-behavior constraints: By aggregating upstream behavior information, providing context for the current behavior, constraining the interactions between behaviors, and ensuring that the input representation of the current behavior can reflect the influence of its upstream behavior; Imposing behavioral constraints: When modeling current behavior, limit the inclusion of downstream behavior information to capture complex user behavior patterns and prevent information leakage; Post-behavior constraints: Design decoupled outputs for each behavior, partition the solution space, and improve model performance through joint optimization.

4. The recommendation method based on multi-behavior information mining according to any one of claims 1 to 3, characterized in that: In step S2, the implementation mechanism of the combined optimization graph convolutional network includes: Utilize the graph convolutional network (GCN) paradigm to effectively simulate high-order connections between users and items to fuse multiple behavioral information; Adjust the adjacency matrix of different behaviors to adapt to the requirements of graph convolution and achieve effective information transfer between behaviors; Capture the interaction information between users and items in each behavior through graph convolution to achieve behavior constraints; In each layer of the graph convolutional network, information about the current behavior and its upstream behavior is combined to perform message propagation and aggregation to implement behavior constraints; Define a hierarchical information transfer mechanism between behaviors, combining the high-level representation of upstream behaviors with the initial representation of the current behavior as input to implement pre-behavior constraints; The representation of each behavior is output independently to meet the needs of multi-task learning, and rich relational representations are obtained by stacking the outputs of different layers, following the post-behavior constraints.

5. The recommendation method based on multi-behavior information mining according to any one of claims 1 to 4, characterized in that: In step S3, the distribution fitting multi-expert network DFME includes the following contrastive learning mechanism: Utilize contrastive learning to adaptively learn the distribution similarity between target and auxiliary behaviors to mitigate the distribution bias between different data sources; Applying the temperature hyperparameter of the softmax function and the similarity calculation function, comparative learning is performed on the user and item sides to generate behavior-specific experts; The behavior representation is decoupled using the Hadamard product operation to prevent gradient conflicts caused by coupled inputs, thereby generating multiple behavior-specific experts.

6. The recommendation method based on multi-behavior information mining according to any one of claims 1 to 5, characterized in that: In step S3, the distribution fitting multi-expert network DFME includes the following behavior fitting expert generation mechanism: Define a dedicated fitting expert for each task to solve the problem of feature distribution deviation, and its output is used for subsequent aggregation process; By controlling the scaling factor of the behavior representation, fine-tuning of the representation space is achieved; Use graph convolution operations to capture the effective information contained in the behavior interaction matrix to enhance the extraction and representation of behavior features; Use graph convolutional networks to process behavior representations and achieve effective information transfer between behaviors through layer-specific parameters; The final expert representation is generated through the Hadamard product operation, similar to the behavior-specific expert, to achieve behavior fitting.

7. The recommendation method based on multi-behavior information mining according to any one of claims 1 to 6, characterized in that: In step S3, the distribution fitting multi-expert network DFME includes the following task aggregation mechanisms: Define the gating of the task, use the feature transformation matrix and bias matrix, and the attention vector as selectors, and calculate the weighted sum of all experts; Utilize behavior-aware graph convolution operations and representation scaling mechanisms to capture the effective components of other behaviors to alleviate the negative transfer caused by feature distribution differences in gated aggregation; By refining the representation and the representation of the target behavior, target aggregation is performed while ensuring that the parameters of the target task are not affected by the gradient update of the auxiliary task and preventing interference from the auxiliary behavior; Apply a stop gradient operation to ensure that the auxiliary behavior does not negatively impact the prediction of the target behavior; The final prediction of the task is formed by weighted summing the outputs of all behaviors; The BPR loss optimization model is used to enhance the recommendation accuracy of the model by comparing the predicted values ​​of positive and negative samples.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the recommendation method based on multi-behavior information mining according to any one of claims 1 to 7 is implemented.

9. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the recommendation method based on multi-behavior information mining according to any one of claims 1 to 7 is implemented.

Images