User intention dissociation multi-behavior recommendation method, system, storage medium and device
Through user intention dissociation and double-layer graph convolution network, the problems of unclear motivation and insufficient high-order relationship modeling in multi-behavior recommendation are solved, accurate recommendation and negative migration mitigation are achieved, and high-order behavior specific relationship modeling is enhanced. , captures common relationships between different behaviors.
Patent Information
- Application Number
- CN202411123977.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-16
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2044-08-16
AI Technical Summary
The existing multi-behavior recommendation method is unclear in the motivation and insufficient modeling of higher-order relationships under the inferred user behavior, resulting in negative migration problems and loss of higher-order information. The comparison learning method has semantic differences for users with weak behavioral associations, confusing the truth. motivation.
Using the user intention dissociation method, the degree of intention attention of users and items under different behaviors is learned through the gated network and the graph convolution network, the specific behavior embedding is calculated, and the enhancement and comparison learning is performed, and the double-layer graph convolution network is enhanced. Advanced relationship modeling, designing self-supervised signals to promote knowledge transfer between multiple behaviors.
Effectively distinguish motivations under different behaviors, alleviate negative migration problems, enhance modeling of specific relationships of higher-order behaviors, capture common relationships between different behaviors, and realize accurate calculation and recommendation of user preference scores.
Smart Images

Figure CN118643222B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of deep learning recommendation technology, and in particular to a user intention dissociation multi-behavior recommendation method, system, storage medium and device. Background Art
[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.
[0003] Recommender systems are widely used in various online applications (e.g., e-commerce, short videos, and social media) to provide more accurate and personalized services. Collaborative filtering (CF) is one of the most successful and effective techniques in recommender systems, which learns user preferences through historical interaction data. The core idea of CF is to map users and items into a latent space to capture latent features or characteristics, and predict the user's preference for a specific item based on the similarity of these features. Early CF models mainly used matrix factorization (MF). However, as linear models, these models have difficulty learning nonlinear relationships between users and items. With the development of deep learning, models based on deep neural networks (DNNs) have replaced MF with nonlinear neural networks to capture complex interaction patterns. In addition, benefiting from the advantages of graph convolutional networks (GCNs) in processing graph-structured data, many studies in recent years have incorporated GCNs into models to explore high-order relationships between users and items, thereby enhancing the representation ability of the model.
[0004] Currently, most traditional recommendation systems use only a single user behavior (such as purchase behavior in e-commerce) to model user preferences. However, in many practical scenarios, there are often multiple interactive relationships (such as click behavior, collection behavior, purchase behavior, etc.), which can more comprehensively reveal the interests and preferences of users. Therefore, a large amount of research is devoted to the development of multi-behavior recommendation models, aiming to utilize rich collaborative signals under multiple behaviors to promote the recommendation of target behaviors. Despite significant progress, there are still three limitations that need to be addressed:
[0005] (1) Existing multi-behavior methods adopt a holistic approach to model user and item representations, ignoring the complex intentions behind the behaviors (such as personal interests, quality, and practicality). In fact, users focus on different items at different stages of their behavior. For example, when clicking, they pay more attention to items that satisfy personal interests. However, when purchasing, they pay more attention to the practicality and quality of the items. Previous methods have encountered the problem of confusion of intentions and failed to accurately infer the true motivation behind each behavior, which may cause irrelevant interactions under auxiliary behaviors to be mistakenly transferred to purchasing behaviors, causing negative transfer problems.
[0006] (2) Existing multi-behavior recommendation methods based on graph convolutional networks (GCNs) have deficiencies in modeling high-order relationships. In order to emphasize the differences between different behaviors, some methods only model specific behavior relationships in independent behavior subgraphs, inadvertently destroying the high-order heterogeneous connections existing in the global heterogeneous graph, resulting in the loss of useful high-order information; in addition, some methods only consider high-order heterogeneous relationships and ignore the modeling of high-order specific behavior relationships, thus failing to accurately capture the unique interaction patterns of users under each behavior.
[0007] (3) Existing multi-behavior recommendation methods based on contrastive learning treat different behavior views of the same user as positive pairs, aiming to achieve knowledge sharing and transfer between multiple behaviors, thereby achieving information enhancement. However, for users with weak correlation between behaviors, there are significant semantic differences between different behaviors. Directly aligning different specific behavior embeddings of users will destroy the uniqueness between multiple behaviors, thereby confusing the real motivations behind different behaviors. Summary of the invention
[0008] The present invention provides a user intention dissociated multi-behavior recommendation method, system, storage medium and device, which effectively solves the problems of unclear motivation of users in inferring different behaviors and insufficient high-order relationship modeling in existing models, and achieves better results in actual scenarios.
[0009] In order to achieve the above object, the present invention adopts the following technical solution:
[0010] A first aspect of the present invention provides a method for dissociating multiple behaviors from user intentions.
[0011] A user intention dissociation multi-behavior recommendation method, comprising:
[0012] Obtain user information, item information, user behavior, and item behavior;
[0013] Based on user information and item information, user intention and item intention are separated to obtain multiple user intention embeddings and multiple item intention embeddings; gating networks are created for user behavior and item behavior respectively to learn the degree of intention attention under different behaviors, obtain different user intention weights and different item intention weights, and input multiple user intention embeddings and multiple item intention embeddings into the gating network to calculate the specific behavior embeddings of users under different behaviors and the specific behavior embeddings of items under different behaviors; user information and item information are converted into overall user embedding and overall item embedding;
[0014] The overall embedding of the user, the overall embedding of the item, the specific behavior embedding of the user under different behaviors, and the specific behavior embedding of the item under different behaviors are enhanced respectively; the enhanced overall embedding of the user is integrated into the enhanced specific behavior embedding of the user under different behaviors to obtain the enhanced specific behavior embedding of the user under different behaviors; the overall embedding of the item is integrated into the enhanced specific behavior embedding of the item under different behaviors to obtain the enhanced specific behavior embedding of the item under different behaviors;
[0015] Calculate the inner product of the enhanced specific behavior embedding of the user under different behaviors and the enhanced specific behavior embedding of the item under different behaviors to obtain the predicted recommendation score for supervised learning; adopt contrastive learning, use the enhanced overall user embedding as an anchor point, and conduct contrastive learning with the enhanced specific behavior embedding of the user under different behaviors; use the enhanced overall item embedding as an anchor point, and conduct contrastive learning with the enhanced specific behavior embedding of the item under different behaviors for self-supervised learning;
[0016] Calculate the inner product of the enhanced specific behavior embedding of the user under the target behavior and the enhanced specific behavior embedding of the item under the target behavior to get the recommendation score, sort the recommendation scores from high to low, select several items with high rankings, and recommend them to the user.
[0017] Furthermore, the method of dissociating user intention and item intention based on user information and item information to obtain multiple user intention embeddings and multiple item intention embeddings includes: projecting user information into different intention spaces to obtain multiple user intention embeddings, wherein the user intention embeddings include a first appearance intention, a first quality intention, and a first brand intention; projecting item information into different intention spaces to obtain multiple item intention embeddings, wherein the item intention embeddings include a second appearance intention, a second quality intention, and a second brand intention.
[0018] Furthermore, the specific behavior embeddings of users under different behaviors and the specific behavior embeddings of items under different behaviors are calculated, and the method includes: under different behaviors, according to the first appearance intention, the first quality intention, the first brand intention, the first appearance intention weight, the first quality intention weight and the first brand intention weight, the specific behavior embeddings of users under different behaviors are calculated; under different behaviors, according to the second appearance intention, the second quality intention, the second brand intention, the second appearance intention weight, the second quality intention weight and the second brand intention weight, the specific behavior embeddings of items under different behaviors are calculated; wherein, the different intention weights of users include: the first appearance intention weight, the first quality intention weight and the first brand intention weight; the different intention weights of items include: the second appearance intention weight, the second quality intention weight and the second brand intention weight.
[0019] Furthermore, the overall embedding of the user, the overall embedding of the item, the specific behavior embedding of the user under different behaviors, and the specific behavior embedding of the item under different behaviors are enhanced respectively, and the method includes: inputting the specific behavior embedding of the user under different behaviors and the specific behavior embedding of the item under different behaviors into the behavior subgraph respectively, performing a graph convolution network, and obtaining enhanced specific behavior embedding of the user under different behaviors and enhanced specific behavior embedding of the item under different behaviors; inputting the overall embedding of the user and the overall embedding of the item into the global heterogeneous graph respectively, performing a graph convolution network, and obtaining enhanced overall embedding of the user and enhanced overall embedding of the item.
[0020] Furthermore, the enhanced overall user embedding is integrated into the enhanced specific behavior embedding of the user under different behaviors to obtain the enhanced specific behavior embedding of the user under different behaviors, and the overall item embedding is integrated into the enhanced specific behavior embedding of the item under different behaviors to obtain the enhanced specific behavior embedding of the item under different behaviors. The method includes: using the attention mechanism to dynamically adjust the importance of the enhanced overall user embedding relative to the enhanced specific behavior embedding of the user under different behaviors, and to adjust the importance of the enhanced overall item embedding relative to the enhanced specific behavior embedding of the item under different behaviors to obtain the propagation weight of the user under different behaviors and the propagation weight of the item under different behaviors; calculating the sum of the specific behavior embedding of the user under different behaviors and the product of the propagation weight of the user under different behaviors and the overall embedding of the user to obtain the enhanced specific behavior embedding of the user under different behaviors; calculating the sum of the specific behavior embedding of the item under different behaviors and the product of the propagation weight of the item under different behaviors and the overall embedding of the item to obtain the enhanced specific behavior embedding of the item under different behaviors.
[0021] Furthermore, a predicted recommendation score is obtained and supervised learning is performed, wherein the method includes: adopting a loss function to reduce the difference between the predicted recommendation score and the actual recommendation score.
[0022] Furthermore, the behavior includes clicking behavior, collecting behavior and purchasing behavior.
[0023] A second aspect of the present invention provides a user intention dissociation multi-behavior recommendation system.
[0024] A user intention dissociation multi-behavior recommendation system, comprising:
[0025] A data acquisition module, which is configured to: acquire user information, item information, user behavior and item behavior;
[0026] The intention-aware multi-behavior modeling module is configured to: dissociate user intention and item intention according to user information and item information to obtain multiple user intention embeddings and multiple item intention embeddings; create gating networks for user behavior and item behavior respectively, learn the degree of intention attention under different behaviors, obtain different user intention weights and different item intention weights, and input multiple user intention embeddings and multiple item intention embeddings into the gating network to calculate the specific behavior embeddings of users under different behaviors and the specific behavior embeddings of items under different behaviors; convert user information and item information into overall user embedding and overall item embedding;
[0027] The enhancement module is configured to: enhance the overall embedding of the user, the overall embedding of the item, the specific behavior embedding of the user under different behaviors, and the specific behavior embedding of the item under different behaviors respectively; integrate the enhanced overall embedding of the user into the enhanced specific behavior embedding of the user under different behaviors to obtain the enhanced specific behavior embedding of the user under different behaviors, integrate the overall embedding of the item into the enhanced specific behavior embedding of the item under different behaviors to obtain the enhanced specific behavior embedding of the item under different behaviors;
[0028] The multi-task learning module is configured to: calculate the inner product of the enhanced specific behavior embedding of the user under different behaviors and the enhanced specific behavior embedding of the item under different behaviors to obtain the predicted recommendation score for supervised learning; adopt contrastive learning to use the enhanced overall user embedding as an anchor point for contrastive learning with the enhanced specific behavior embedding of the user under different behaviors; use the enhanced overall item embedding as an anchor point for contrastive learning with the enhanced specific behavior embedding of the item under different behaviors for self-supervised learning;
[0029] The system recommendation module is configured to: calculate the inner product of the enhanced specific behavior embedding of the user under the target behavior and the enhanced specific behavior embedding of the item under the target behavior, obtain the recommendation score, arrange the recommendation scores from high to low, select several items with high rankings, and recommend them to the user.
[0030] A third aspect of the present invention provides a computer-readable storage medium.
[0031] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps in the user intention dissociation multi-behavior recommendation method as described in the first aspect above.
[0032] A fourth aspect of the present invention provides a computer device.
[0033] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the method for recommending multiple behaviors by dissociating user intentions as described in the first aspect above are implemented.
[0034] Compared with the prior art, the present invention has the following beneficial effects:
[0035] The present invention first incorporates the influence of intention into multi-behavior semantic modeling to infer the user's true motivation, and designs a two-layer graph convolutional network to enhance the model's high-order relationship modeling capabilities. Finally, the model is optimized through a multi-task learning paradigm to obtain the feature vectors of users and items, and the inner product method is used to calculate the user's preference score for the item, achieving accurate recommendations for users.
[0036] The present invention decouples user intentions and integrates the impact of intentions into multi-behavior modeling, which can distinguish the motivations under different behaviors and effectively alleviate the negative transfer problem.
[0037] The present invention utilizes rich high-order heterogeneous relationships as supplementary information to strengthen the modeling of high-order behavior-specific relationships and effectively alleviate the problem of single behavior data sparsity.
[0038] The present invention introduces a novel contrastive learning task and designs additional self-supervision signals to promote the transfer of shared knowledge between multiple behaviors and effectively capture the common relationships between different behaviors. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.
[0040] Figure 1 Schematic diagram of the architecture of the user intention dissociation multi-behavior recommendation system of the present invention;
[0041] Figure 2 It is a detailed schematic diagram of the model of module 1 of the present invention;
[0042] Figure 3 It is a detailed schematic diagram of the model of module 2 of the present invention;
[0043] Figure 4 It is a schematic diagram of multi-behavior supervised learning in the model of module three of the present invention;
[0044] Figure 5 It is a schematic diagram of behavior perception contrast learning in the model of module three of the present invention. DETAILED DESCRIPTION
[0045] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0046] In this embodiment, a known shopping platform dataset is selected as the operation object, and implicit user feedback is used, which is closer to the actual recommendation scenario. The purchase behavior is regarded as the target behavior, and the click and collection behaviors are regarded as auxiliary behaviors to verify the effectiveness and scalability of the present invention on a large-scale e-commerce platform. It should be noted that the known shopping platform dataset is authorized and allowed to be publicly used data, which is obtained through legal channels.
[0047] Embodiment 1
[0048] This embodiment provides a user intention dissociation multi-behavior recommendation method, comprising: obtaining user behavior and items related to the user behavior, using a recommendation model, and obtaining recommended items; wherein the recommendation model includes an intention-aware multi-behavior modeling module, an enhancement module, and a multi-task recommendation module;
[0049] 1. Use the intention-aware multi-behavior modeling module to separate the multiple intentions of users under different behaviors, and model the degree to which users are affected by intentions under various behaviors, so as to more accurately understand the motivations and preferences of users under different behaviors. The specific steps include:
[0050] (1) Initializing embedding parameters: Initialize embedding parameters for users and items as their corresponding feature vectors to reflect the corresponding feature information.
[0051] (2) Intent dissociation: Using dissociated representation learning technology, the user’s multiple intentions under different behaviors are separated from the overall representation. Each user is embedded with multiple intents, where different intents represent different purposes of the user’s interaction with the item and are independent of each other.
[0052] (3) Design of Intent Fusion Gated Network (IFGN): This network uses a gating mechanism to model the influence of different intentions of users under various behaviors, and fuses multiple intentions of users according to their importance to represent the features of users under specific behaviors.
[0053] 2. Use enhancement modules to fully utilize the structural information under various behaviors. To this end, a two-layer graph convolutional network is used to model high-order relationships from the perspectives of global behavior and specific behavior. The specific steps include:
[0054] (1) Mapping: Constructing a global heterogeneous graph using the user-item interaction matrix , to model high-order heterogeneous relationships. Including user and item nodes, as well as user-item interaction edges under various behaviors. And the heterogeneous graph Split into K Behavior subgraph , to model higher-order behavior-specific relationships.
[0055] (2) Behavior-specific message propagation: Message propagation is performed on each behavior subgraph to learn the behavior-specific embeddings of nodes. By propagating messages between neighboring nodes in the behavior subgraph, the embeddings of users and items are iteratively updated to model high-order behavior-specific relationships.
[0056] (3) Global Heterogeneous Message Propagation: Message propagation is performed on a global heterogeneous graph to learn globally-aware embeddings of nodes. By performing message propagation between neighbor nodes in a global heterogeneous graph, high-order heterogeneous connections are injected into the globally-aware embeddings to model behavior-specific relationships.
[0057] (4) Adaptive information fusion: The global perception embedding is adaptively fused into the specific behavior embedding using the attention mechanism, thereby enhancing the expressive power of the specific behavior embedding.
[0058] 3. Use the multi-task learning module to design a multi-task learning framework to optimize model performance and achieve knowledge transfer between different tasks. The specific steps include:
[0059] (1) Multi-behavior supervised learning tasks: Utilize the supervisory signals under different behaviors and use non-sampled learning loss as the training objective to optimize the model. This method avoids the noise and bias problems caused by traditional sampling methods by directly utilizing all user-item interaction data.
[0060] (2) Contrastive learning task: The user’s global heterogeneous perceptual embedding is used as the anchor, the embeddings of different views of the same user are used as positive pairs, and the embeddings of other users are used as negative pairs to achieve the transfer of common features, thereby capturing the commonalities between multiple behaviors and avoiding the semantic confusion problem caused by directly aligning the embeddings of different specific behaviors of users.
[0061] Embodiment 2
[0062] like Figure 1 As shown, the user intention dissociation multi-behavior recommendation system described in this embodiment includes: a data acquisition module, an intention-aware multi-behavior modeling module, etc.
[0063] 1. Data Acquisition Module
[0064] Used to obtain user information, item information, user behavior, and item behavior.
[0065] 2. Module 1 (Intention-aware multi-behavior modeling module), such as Figure 2 As shown:
[0066] 2.1 Embedding Vector Initialization
[0067] In order to describe the characteristics of users and items, the corresponding embedding vectors are initialized according to the IDs of each user and item. and In addition, for each behavior , initializes an embedding vector , which aims to distinguish different types of interactions between users and items (e.g., click, add to cart, and purchase).
[0068] 2.2 Intention Dissociation Layer
[0069] Since user behavior is usually driven by multiple intentions, this model uses dissociative representation learning technology to project the embeddings of users and items into In the intention space, the dissociation operation is realized. Specifically, for the node (Representing the user or items ), its corresponding intention Obtained through:
[0070] ; (1)
[0071] in, Representation Node Intention The embedding representation of and is a learnable parameter, It is a non-linear activation function.
[0072] Finally, each user and item gets a set of intent embeddings , .
[0073] 2.3 Intent Fusion Gating Network
[0074] In order to accurately model the degree of attention users pay to different intentions at different behavior stages, this model designs an intention fusion gating network, which includes two key steps. First, an independent gating network is designed for each behavior to quantify the fusion weight of the intent to represent the relative driving degree of multiple intentions under the behavior. The specific operations are as follows:
[0075] ; (2)
[0076] in, represents the intention weight vector, indicating the user In behavior The relative driving forces of all intentions below; Indicates user The embedding vector of Expressing behavior The embedding vector of are the learnable parameters of behavior k, Represents the bitwise multiplication of elements (Hadamard product).
[0077] Second, based on the corresponding behavior weights, the intent embeddings are fused to generate specific behavior embeddings, aiming to encode unique behavior semantics:
[0078] ; (3)
[0079] in, Indicates user In behavior Specific behavior embedding under is the intention weight vector No. elements, representing the user In behavior Under the intention the extent of the impact; Indicates user Intention Embedding.
[0080] Similarly, the specific behavior embedding of items is calculated in the same way. Finally, IFGN generates The embedding of specific behaviors is represented as and .
[0081] 3. Module 2 (enhanced module), such as Figure 3 As shown:
[0082] 3.1 Specific behavior message dissemination
[0083] In order to accurately capture the user's specific behavioral interaction pattern, Message propagation is performed on the network, and high-order connections (for example, when the information is transmitted for the first time, the neighbor information is propagated to itself to find out who bought which item; when the information is transmitted for the second time, the neighbor's neighbor information is propagated to itself to find out that two users bought the same item) are injected into the behavior-specific embedding, thereby modeling high-order specific behavior relationships. Specifically, for behavior , the embedding of users at different propagation layers is calculated as follows:
[0084] ; (4)
[0085] in, Indicates user In behavior Next Layer-propagated specific behavior embedding, Indicates items In behavior Next Layer-specific behavior embedding, and The users are and items In the subgraph The set of neighbors in is the transformation matrix for a particular layer, Expressing behavior In the Embedding of layers, via layer-specific transformation matrices renew:
[0086] ; (5)
[0087] It should be noted that , , Similarly, for items, the behavior of each layer can be calculated in a similar way: Representation.
[0088] 3.2 Global Heterogeneous Message Propagation
[0089] Next, in order to further utilize the rich heterogeneous high-order collaboration signals in multi-behavior interaction data. We implement global heterogeneous message propagation on the CNN to inject high-order heterogeneous connections into the global-aware embedding for modeling high-order heterogeneous relations, aiming to capture the global behavioral interaction patterns of users. Specifically, in order to distinguish the types of interactive behaviors between users and items, we introduce the embedding of behaviors into the message passing paradigm:
[0090] ; (6)
[0091] in, Indicates user In the The global-aware embedding of the layer, Indicates items in front The global-aware embedding of the layer, and It is a globally heterogeneous graph Medium User and items The set of neighbors.
[0092] It is worth noting that Similarly, for items, the global-aware embeddings of each layer can be obtained in a similar way.
[0093] 3.3 Adaptive Information Fusion
[0094] go through After the layer message propagation, at each layer, the node (user or items )get Behavior-specific embedding , and a globally aware embedding In addition, for each behavior, the corresponding embedding Since embeddings at different layers represent relationships of different orders, the present invention aggregates the embeddings obtained at each layer into the final embedding by an average operation:
[0095] ; (7)
[0096] Behavior-specific embedding Learn the unique interaction characteristics of users (items) in specific behaviors. In addition, global perception embedding The global behavioral interaction patterns of users (items) are encoded from a global perspective, aiming to learn their overall preferences under different behaviors.
[0097] Next, in order to synergistically utilize the contextual information of specific behaviors and global behaviors, the behavior-specific embedding is combined with the global-aware embedding to obtain the final behavior-specific embedding and enhance the expressiveness of the behavior-specific embedding. However, since the degree to which users' interaction preferences under different behaviors depend on the global behavior interaction pattern is different, the attention mechanism is used to dynamically adjust the importance of the global-aware embedding relative to the behavior-specific embedding:
[0098] ; (8)
[0099] in, is the attention weight, through the intermediate factor The SoftMax operation calculates the attention weight; and is the transformation matrix;
[0100] Finally, behavior-specific embeddings According to the attention weight Updated to:
[0101] ; (9)
[0102] Module 3 (Multi-task learning module)
[0103] 4.1 Multiple behavior supervised learning tasks, such as Figure 4 As shown:
[0104] In order to improve the stability of model parameter learning, non-sampling learning loss is used as the training target to optimize the model of the present invention to alleviate the problems related to sampling noise and sample bias in sampling-based methods. The definition is as follows:
[0105] ; (10)
[0106] in, Indicates user In behavior The collection of items that have been interacted with. and Respectively express behavior Next item The positive and negative weights of represents the dimension of embedding, represents the mth number of embeddings of behavior k, represents the nth number of embeddings of behavior k, represents the mth number of specific behavior embeddings of user u under specific behavior k, represents the nth number of specific behavior embeddings of user u under specific behavior k, represents the mth number of specific behavior embeddings of user i under specific behavior k, represents the nth number of specific behavior embeddings of user i under specific behavior k. Is a user With items In behavior The predicted interaction probability under:
[0107] ; (11)
[0108] in, represents a diagonal matrix whose diagonal elements are equal to .
[0109] 4.2 Contrastive learning tasks for behavior perception, such as Figure 5 As shown:
[0110] Although users show different interaction preferences under different behaviors, these behaviors usually have certain commonalities. Since the global-aware embedding learned in the context of global behaviors can capture the global interaction patterns of users and reflect the overall preferences of users, a contrastive learning paradigm is designed in which the global-aware embedding is used as an anchor to compare with the specific behavior embedding to achieve knowledge transfer of shared features. Specifically, the embedding of the same user is are regarded as positive samples, and the embedding pairs from different users is regarded as a negative sample. InfoNCE loss is used to enhance the consistency of different perspectives of the same user to distinguish the behavioral preferences between different users:
[0111] ; (12)
[0112] in, is the temperature coefficient of contrast learning, represents the inner product of two vectors and is used to calculate the similarity between embeddings. Indicates user The specific behavior of k embedding, represents the specific behavior k embedding of user u.
[0113] Similarly, the contrast loss of the items is calculated Finally, the loss functions of users and items under all behaviors are combined to obtain the final loss function :
[0114] ; (13)
[0115] 4.3 Joint Optimization
[0116] After defining all tasks, all loss functions are combined to get a total loss function , as shown below:
[0117] ; (14)
[0118] in, is the number of behavior types, represents the loss coefficient of the supervised learning task associated with each behavior, satisfying ; represents the loss coefficient of the behavior perception contrast learning task. In addition, in order to prevent overfitting, the parameter application Regularization, where is the regularization coefficient.
[0119] In order to verify the rationality and performance of the present invention, experiments are conducted on two datasets, and two indicators are used to compare with existing benchmark models, including a single behavior model and a multi-behavior model.
[0120] HR@K: Measures whether there are positive samples successfully recommended to the user (i.e., the score ranks in the top K). If yes, it is counted as 1, otherwise it is 0.
[0121] NDCG@K: Normalized Discounted Cumulative Gain, used to measure the relevance of recommendation results.
[0122] Table 1 shows the results of the two indicators on the test set, with K set to 10 and 50.
[0123]
[0124] As can be seen from Table 1, the method of the present invention significantly outperforms the existing baseline models on both datasets. This significant performance improvement is mainly attributed to the three key innovations of the present invention: (1) Decoupling user intentions and incorporating the impact of intentions into multi-behavior modeling can distinguish the motivations under different behaviors and effectively alleviate the negative transfer problem. (2) Using rich high-order heterogeneous relationships as supplementary information, the modeling of high-order behavior-specific relationships is strengthened, effectively alleviating the problem of single behavior data sparsity. (3) Introducing a novel contrastive learning task and designing additional self-supervisory signals to promote the transfer of shared knowledge between multiple behaviors and effectively capture the common relationships between different behaviors.
[0125] Embodiment 3
[0126] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in the method for recommending multiple behaviors by dissociating user intentions as described in the first embodiment are implemented.
[0127] Embodiment 4
[0128] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the method for recommending multiple behaviors by dissociating user intentions as described in the first embodiment are implemented.
[0129] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A user intention dissociation multi-behavior recommendation method, characterized in that: include: Obtain user information, item information, user behavior, and item behavior; Based on user information and item information, user intention and item intention are separated to obtain multiple user intention embeddings and multiple item intention embeddings; gating networks are created for user behavior and item behavior respectively to learn the degree of intention attention under different behaviors, obtain different user intention weights and different item intention weights, and input multiple user intention embeddings and multiple item intention embeddings into the gating network to calculate the specific behavior embeddings of users under different behaviors and the specific behavior embeddings of items under different behaviors; user information and item information are converted into overall user embedding and overall item embedding; Enhance the overall embedding of users, the overall embedding of items, the specific behavior embedding of users under different behaviors, and the specific behavior embedding of items under different behaviors respectively; The enhanced user overall embedding is integrated into the enhanced user specific behavior embedding under different behaviors to obtain the enhanced specific behavior embedding of the user under different behaviors; the item overall embedding is integrated into the enhanced item specific behavior embedding under different behaviors to obtain the enhanced specific behavior embedding of the item under different behaviors; Calculate the inner product of the enhanced specific behavior embedding of users under different behaviors and the enhanced specific behavior embedding of items under different behaviors to obtain the predicted recommendation score for supervised learning; Using contrastive learning, the enhanced overall user embedding is used as an anchor point to perform contrastive learning with the enhanced specific behavior embedding of the user under different behaviors; The enhanced overall embedding of the item is used as an anchor point to compare with the enhanced specific behavior embedding of the item under different behaviors for self-supervised learning; Calculate the inner product of the enhanced specific behavior embedding of the user under the target behavior and the enhanced specific behavior embedding of the item under the target behavior to get the recommendation score, sort the recommendation scores from high to low, select several items with high rankings, and recommend them to the user.
2. The user intention dissociation multi-behavior recommendation method according to claim 1, characterized in that: The method of dissociating user intention and item intention based on user information and item information to obtain multiple user intention embeddings and multiple item intention embeddings includes: projecting user information into different intention spaces to obtain multiple user intention embeddings, wherein the user intention embeddings include a first appearance intention, a first quality intention, and a first brand intention; projecting item information into different intention spaces to obtain multiple item intention embeddings, wherein the item intention embeddings include a second appearance intention, a second quality intention, and a second brand intention.
3. The user intention dissociation multi-behavior recommendation method according to claim 2, characterized in that: Calculate the specific behavior embedding of users under different behaviors and the specific behavior embedding of items under different behaviors. The method process includes: under different behaviors, according to the first appearance intention, the first quality intention, the first brand intention, the first appearance intention weight, the first quality intention weight and the first brand intention weight, calculate the specific behavior embedding of users under different behaviors; under different behaviors, according to the second appearance intention, the second quality intention, the second brand intention, the second appearance intention weight, the second quality intention weight and the second brand intention weight, calculate the specific behavior embedding of items under different behaviors; wherein, the different intention weights of users include: the first appearance intention weight, the first quality intention weight and the first brand intention weight; the different intention weights of items include: the second appearance intention weight, the second quality intention weight and the second brand intention weight.
4. The user intention dissociation multi-behavior recommendation method according to claim 1, characterized in that: The overall embedding of the user, the overall embedding of the item, the specific behavior embedding of the user under different behaviors, and the specific behavior embedding of the item under different behaviors are enhanced respectively. The method includes: inputting the specific behavior embedding of the user under different behaviors and the specific behavior embedding of the item under different behaviors into the behavior subgraph respectively, performing a graph convolution network, and obtaining enhanced specific behavior embedding of the user under different behaviors and the specific behavior embedding of the item under different behaviors; inputting the overall embedding of the user and the overall embedding of the item into the global heterogeneous graph respectively, performing a graph convolution network, and obtaining enhanced overall embedding of the user and enhanced overall embedding of the item.
5. The user intention dissociation multi-behavior recommendation method according to claim 1, characterized in that: The enhanced overall user embedding is integrated into the enhanced specific behavior embedding of the user under different behaviors to obtain the enhanced specific behavior embedding of the user under different behaviors, and the overall item embedding is integrated into the enhanced specific behavior embedding of the item under different behaviors to obtain the enhanced specific behavior embedding of the item under different behaviors. The method includes: using the attention mechanism to dynamically adjust the importance of the enhanced overall user embedding relative to the enhanced specific behavior embedding of the user under different behaviors, and to adjust the importance of the enhanced overall item embedding relative to the enhanced specific behavior embedding of the item under different behaviors to obtain the propagation weight of the user under different behaviors and the propagation weight of the item under different behaviors; calculating the sum of the specific behavior embedding of the user under different behaviors and the product of the propagation weight of the user under different behaviors and the overall embedding of the user to obtain the enhanced specific behavior embedding of the user under different behaviors; calculating the sum of the specific behavior embedding of the item under different behaviors and the product of the propagation weight of the item under different behaviors and the overall embedding of the item to obtain the enhanced specific behavior embedding of the item under different behaviors.
6. The user intention dissociation multi-behavior recommendation method according to claim 1, characterized in that: The predicted recommendation score is obtained and supervised learning is performed, wherein the method includes: adopting a loss function to reduce the difference between the predicted recommendation score and the actual recommendation score.
7. The user intention dissociation multi-behavior recommendation method according to any one of claims 1 to 6, characterized in that: The behaviors include clicking behaviors, collecting behaviors and purchasing behaviors.
8. User intention dissociation multi-behavior recommendation system, characterized by: include: A data acquisition module, which is configured to: acquire user information, item information, user behavior and item behavior; The intention-aware multi-behavior modeling module is configured to: dissociate user intention and item intention according to user information and item information to obtain multiple user intention embeddings and multiple item intention embeddings; create gating networks for user behavior and item behavior respectively, learn the degree of intention attention under different behaviors, obtain different user intention weights and different item intention weights, and input multiple user intention embeddings and multiple item intention embeddings into the gating network to calculate the specific behavior embeddings of users under different behaviors and the specific behavior embeddings of items under different behaviors; convert user information and item information into overall user embedding and overall item embedding; An enhancement module is configured to: respectively enhance the overall embedding of the user, the overall embedding of the item, the specific behavior embedding of the user under different behaviors, and the specific behavior embedding of the item under different behaviors; The enhanced user overall embedding is integrated into the enhanced user specific behavior embedding under different behaviors to obtain the enhanced specific behavior embedding of the user under different behaviors; the item overall embedding is integrated into the enhanced item specific behavior embedding under different behaviors to obtain the enhanced specific behavior embedding of the item under different behaviors; A multi-task learning module is configured to: calculate the inner product of the enhanced specific behavior embedding of the user under different behaviors and the enhanced specific behavior embedding of the item under different behaviors, obtain the predicted recommendation score, and perform supervised learning; Using contrastive learning, the enhanced overall user embedding is used as an anchor point to perform contrastive learning with the enhanced specific behavior embedding of the user under different behaviors; The enhanced overall embedding of the item is used as an anchor point to compare with the enhanced specific behavior embedding of the item under different behaviors for self-supervised learning; The system recommendation module is configured to: calculate the inner product of the enhanced specific behavior embedding of the user under the target behavior and the enhanced specific behavior embedding of the item under the target behavior, obtain the recommendation score, arrange the recommendation scores from high to low, select several items with high rankings, and recommend them to the user.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the user intention dissociation multi-behavior recommendation method as described in any one of claims 1 to 7 are implemented.
10. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the user intention dissociation multi-behavior recommendation method according to any one of claims 1 to 7 are implemented.
