Recommendation method and electronic device based on user usage intention perception

By constructing hypergraphs and heterogeneous information networks, and combining user intent recognition and item association, the problem of difficult modeling of user intent and item association in recommendation systems is solved, achieving more accurate and personalized recommendation results.

CN120492740BActive Publication Date: 2025-09-09NANJING UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202510984157.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2025-07-17
Publication Date
2025-09-09
Estimated Expiration
2045-07-17

AI Technical Summary

Technical Problem

Existing recommendation systems find it difficult to effectively capture the underlying intentions behind user behavior and the complex semantic associations between items, resulting in recommendation results that do not meet user needs.

Method used

By generating initial embedding vectors for items, performing intent encoding and clustering, building hypergraphs and heterogeneous information networks, and utilizing dual-graph representation learning and attention mechanisms, we can identify user intent and model complex associations between items, and make recommendations based on user historical purchase characteristics.

Benefits of technology

It significantly improves the accuracy and personalization of recommendation results, can more finely identify the potential intentions behind user behavior, and effectively capture high-order semantic associations between items, achieving more accurate personalized recommendations.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120492740B_ABST
    Figure CN120492740B_ABST
Patent Text Reader

Abstract

The present invention provides a recommendation method and electronic device based on user intention perception, wherein the method includes: generating an initial embedding vector; encoding user behavior intention; constructing a dual network; decoding intention based on dual-graph representation learning; intention decoupling; generating user embeddings and recommending items. In terms of recommendation accuracy, the present invention, through the intention decoupling mechanism, can more finely identify the potential intentions behind user behavior, and accordingly model the semantic associations between users and items at the intention level, thereby improving the personalized matching degree of recommendation results. In terms of recommendation relevance, a heterogeneous information network and a hypergraph structure are constructed simultaneously. The former is used to characterize the complex interactions of multiple types of nodes and relationships, while the latter effectively captures high-order semantic associations, thereby realizing multi-level modeling of user-item interaction scenarios, significantly improving the accuracy and relevance of recommendations.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a recommendation method and electronic device based on user usage intention perception. Background Art

[0002] As an efficient information filtering tool, recommendation systems have been frequently used in e-commerce platforms in recent years to help consumers reduce search time and improve shopping satisfaction.

[0003] Current research on recommender systems focuses on improving the accuracy of recommendations, often overlooking the intentions behind user behavior. However, in real-world scenarios, user needs are often context-dependent and heterogeneous. The same user may exhibit different interests and preferences in different scenarios, while different users may have similar behavioral intentions. If recommendation models fail to effectively identify and leverage the underlying intentions behind user behavior, recommendations will not meet users' actual needs. Therefore, effectively capturing and leveraging the underlying intentions in user behavior is a key challenge in current recommender system research.

[0004] In particular, existing graph neural network-based methods mostly use simple homogeneous graph structures to model the interactive relationship between users and items, which makes it difficult to effectively capture the complex high-order associations and diverse semantic relationships in recommendation scenarios. Although some methods have attempted to introduce hypergraph structures to enhance the modeling capabilities of multi-object collaborative relationships, these methods often ignore the important role of item-side information in characterizing the relationship between items and identifying user intentions. In fact, auxiliary information such as item attributes, categories, and text descriptions usually exists in a scattered and redundant form, lacking an effective integration mechanism, making it difficult to fully explore its value. Therefore, how to efficiently integrate multi-source side information of items and build a more expressive item relationship modeling framework on this basis is an important issue that needs to be urgently addressed in the current field of recommendation systems. Summary of the Invention

[0005] As a core tool for assisting users in efficiently screening personalized items, recommendation systems face two key challenges in practical applications: first, user behavior often implies multiple potential usage intentions, such as "for personal use," "gifting," or "price comparison"; second, there are complex semantic associations and high-order interactive relationships between items, which are difficult to effectively model using traditional methods. To address these issues, the present invention proposes a recommendation method and electronic device based on user usage intention perception. This method aims to automatically identify potential intentions from user behavior and collaboratively model the complex association structure between items, thereby significantly improving the accuracy and personalization of recommendation results.

[0006] In one aspect, the present invention provides a recommendation method based on user usage intention perception, comprising the following steps:

[0007] Step 1: Generate the initial embedding vector of the item: Map the ID of the item into the vector space to obtain the initial embedding vector of each item, which are recorded as , where d is the dimension of the embedding vector;

[0008] Step 2: User behavior intention encoding, split the initial embedding vector into K intention sub-vectors, cluster the sub-vectors of all items under each intention, and generate the intention feature vector ;

[0009] Step 3: Build a dual network, including building a hypergraph whose nodes are items and whose hyperedges include co-purchase hyperedges and intention hyperedges; and building a heterogeneous information network whose nodes include items and side information, and whose edges represent predefined relationships between nodes;

[0010] Step 4: Intent decoding based on dual-graph representation learning, that is, adjusting the weight of the node in the hypergraph information propagation through the hypergraph attention network, so that the hyperedge with high semantic and intent matching with the current node is given a greater weight, including node-to-hyperedge attention aggregation and hyperedge-to-node attention aggregation, and updating the intent embedding of each node ; Update the nodes of the heterogeneous information network through the graph convolutional network, and use the triplet scoring mechanism to optimize and obtain the new node embedding Fusion and Generate an embedding of a specific intent; perform intent decoupling classification based on the embedding of the specific intent;

[0011] Step 5: Aggregate the features of the user's historical purchases through the attention mechanism to obtain the user's embedding of intent k;

[0012] Step 6: Calculate the intention relevance score between the user and the item, and generate a Top-N recommendation list by sorting the scores.

[0013] Furthermore, the sub-vectors of all items under each intent are clustered as described in step 2 to generate the intent feature vector Specifically, for the kth potential intent, the intent sub-vectors of all items are processed by clustering algorithm to generate several cluster centers, which are the intent feature vectors. .

[0014] Furthermore, step 3 also includes the following process: filtering intent hyperedges that are not related to intent k. When the similarity score between an item and intent k is greater than a predefined threshold, there is an intent hyperedge between the two, otherwise it does not exist.

[0015] Furthermore, the attention aggregation from the node to the hyperedge in step 4 includes, through the aggregation function Aggregate all hyperedges j Connected node features, update the embedding representation of the hyperedge about intention k

[0016]

[0017] in, Indicates that under the k-th intention, the hyperedge he j The embedding vector of Represents super edge j midpoint The attention weight of

[0018] The hyperedge-to-node attention aggregation includes, through the aggregation function Aggregate all hyperedge information associated with the current node to update the intent embedding of each node,

[0019]

[0020] in, It is a super edge For Node The attention weight of .

[0021] Furthermore, the fusion described in step 4 and Generating intent-specific embeddings involves using a feature fusion function Fusion is performed to generate item embeddings with specific intent. The formula is: .

[0022] Furthermore, in the fusion and After generating the embedding for a specific intent, we also need to learn to align the embedding semantics through cross-network comparison. Specifically, for items , its intention embedding in the hypergraph and embedded in heterogeneous information networks Constitute a positive sample pair , and then randomly select its negative sample pair ,in Finally, the parameters of the InfoNCE loss optimization model are used to ensure that the embeddings generated by the two networks are semantically aligned. The parameters of the InfoNCE loss optimization model are

[0023]

[0024] Among them, Similarity function, is the temperature parameter, For items The set of negative samples across the network.

[0025] Furthermore, the intention decoupling classification based on the embedding of the specific intent in step 4 includes classifying the intent using a classification model, Intent embedding update.

[0026] Furthermore, the calculation of the user-item intention relevance score in step 6 includes using the similarity function Calculating users With all candidate items Relevance score

[0027]

[0028] in, For items Embedding on intent k, Represents a user Embedding on intent k.

[0029] Furthermore, step 7 is included, constructing a joint optimization function to improve the performance of the method, and the loss function of the joint optimization function is the sum of the loss functions in steps 4 and 5.

[0030] In a second aspect of the present invention, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory, wherein when the processor executes the program, the steps of any of the above-mentioned recommendation methods based on user intention perception are implemented.

[0031] In terms of recommendation accuracy, this method, based on user intent perception, uses an intent decoupling mechanism to more precisely identify the underlying intentions behind user behavior and, based on this, model the semantic associations between users and items at the intent level, thereby improving the personalized matching of recommendation results. Regarding recommendation relevance, both a heterogeneous information network and a hypergraph structure are constructed. The former is used to characterize the complex interactions between multiple types of nodes and relationships, while the latter effectively captures high-order semantic associations, thereby achieving multi-level modeling of user-item interaction scenarios and significantly improving the accuracy and relevance of recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0033] Figure 1 This is a flow chart of a recommendation method based on user intent perception according to an embodiment of the present invention. DETAILED DESCRIPTION

[0034] In order to enable those skilled in the art to better understand the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of this application.

[0035] It should be noted that the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or are inherent to these processes, methods, products or apparatus.

[0036] In one aspect, the present invention provides a recommendation method based on user usage intention perception, comprising the following steps:

[0037] Step 1: Generate the initial embedding vector of the item: Map the ID of the item into the vector space to obtain the initial embedding vector of each item, which are recorded as , where d is the dimension of the embedding vector.

[0038] The purpose of this step is to generate an initial representation vector for each item, which serves as the basis for subsequent intent encoding. Specifically, the item ID can be mapped to the vector space through one-hot encoding and linear mapping.

[0039] Step 2: User behavior intention encoding, split the initial embedding vector into K intention sub-vectors, ,in Corresponding to the Kth intent; cluster the sub-vectors of all items under each intent to generate the intent feature vector The purpose of this step is to decompose the initial embedding vector of each item into multiple sub-vectors corresponding to potential usage intentions, so as to accurately represent the motivation of users to interact with items.

[0040] Specifically, it includes two steps: vector slicing and generating intent prototypes:

[0041] S2.1 Vector Slice: For any item , the embedding vector obtained in step 1 Specifically, assuming the number of intents is K, based on this value, Divide into K sub-vectors, formalized as a set containing K intentions Each sub-vector Corresponding items Embedding vector under the kth potential usage intent semantics.

[0042] S2.2 Generate Intent Prototype: In order to identify the behavior pattern of each intent, this step clusters the sub-vectors of the items under each intent. For the kth potential intent, the intent sub-vectors of all items are processed by clustering algorithms, including but not limited to K-means, mean, etc. , generate several cluster centers, which are called intention "prototype vectors" , represents the feature representation of the kth potential intent.

[0043] The set of intention sub-vectors obtained in this step and the intent feature vector , providing support for the dual graph representation learning construction in step 4.

[0044] Step 3: Build a dual network, including building a hypergraph , whose nodes are items, and hyperedges include co-purchase hyperedges and intention beyond the boundary ; and, building heterogeneous information networks , whose nodes contain items and side information, and edges represent predefined relationships between nodes.

[0045] HV is a node set, each node Both indicate an item; is a set of hyperedges, each hyperedge The co-purchase hyperedge connects all items purchased by the same user, reflecting the high-order correlation of items under the co-purchase relationship. The intention hyperedge connects all nodes in the hypergraph, reflecting the characteristics related to user usage intention. In other words, any intention hyperedge It can be regarded as a collection of items, and all items in the collection have hyperedge correlation.

[0046] Preferably, this step filters out hyperedges that are not related to the intent k by a predefined threshold: The similarity score (e.g., cosine value, vector inner product value) with intent k is greater than the threshold When , there is a hyperedge between the two; otherwise, there is no hyperedge.

[0047] Construct the heterogeneous information network based on the side information of the items (such as brand, category, usage category, etc.) VS is a set of nodes (e.g., items, brands, categories, etc.); E is a set of edges, which includes various predefined relationships between nodes (e.g., item-brand, item-category, etc.). When the relationship between nodes satisfies the predefined relationship, an edge exists between the two nodes; otherwise, no edge exists.

[0048] Step 4: Intent decoding based on dual-graph representation learning aims to integrate hypergraph representation learning and graph representation learning to enhance the representation of items related to specific intents. Specifically, it includes:

[0049] S4.1 Hypergraph Representation Learning: The hypergraph attention network HGAN is used to adjust the weight of nodes in the hypergraph information propagation, so that hyperedges with high semantic and intention matching with the current node are given greater weight, including attention aggregation from node to hyperedge and from hyperedge to node, and updating the intention embedding of each node. ;

[0050] Specifically, the attention aggregation from the node to the hyperedge is achieved through an aggregation function (e.g. sum, mean, weighted average) aggregate all hyperedges j Connected node features, update each hyperedge The embedding representation of intent k is shown in formula (1).

[0051] (1)

[0052] in, Indicates that under the k-th intention, the hyperedge The embedding vector of is an aggregation function defined on a hyperedge, Represents the super edge midpoint The attention weight is used to quantify the contribution of the node to the hyperedge feature representation. Considering that the higher the semantic consistency between a node and its hyperedge, the more representative its modeling of the hyperedge is. and hyperedge embedding The similarity of the node determines the attention weight of the node , as shown in formula (2).

[0053] (2)

[0054] in, is a learnable attention vector, represents element-wise product, LeakyReLU is the activation function; For super edge The feature embedding of is obtained by aggregating the features of all nodes and can be expressed as: .

[0055] The attention aggregation from the hyperedge to the node is achieved through an aggregation function Aggregate all hyperedge information associated with a node (e.g., weighted sum, mean, or linear fusion) to update each node The intention embedding can be formalized as formula (3),

[0056] (3)

[0057] in, It is a super edge For Node The attention weight can be obtained by formula (4),

[0058] (4)

[0059] in, is another learnable attention vector under the kth intent. This process considers the correlation between the hyperedge embedding and the collective intent, so that the hyperedge with a high degree of match with the current node semantics and intent is given a greater weight.

[0060] S4.2 Graph Representation Learning: Updating nodes in heterogeneous information networks through graph convolutional networks and optimizing them with triplet scoring mechanisms to obtain new node embeddings ; The purpose is to enhance the correlation between items in terms of side information.

[0061] Specifically, there are multiple triples in heterogeneous information networks, and their set can be recorded as , where h and t are the head entity and tail entity respectively, and r represents the relationship between them. The initial representation of each node and relationship in : , , where m is the characteristic dimension of the relationship. First, by introducing two mapping matrices , mapping the entity vector to the vector space to which the relationship belongs, can be shown by formulas (5) and (6).

[0062] (5)

[0063] (6)

[0064] Then, define a scoring function It is used to evaluate the matching degree between the head entity and the tail entity in a triple under a given relationship. Formula (7) describes the evaluation process under the semantics of each usage intent.

[0065] (7)

[0066] in represents L2 regularization. Finally, the parameters in the node representation learning process are optimized by minimizing the score difference between positive samples (real triplets) and negative samples (error triplets artificially constructed by random sampling), as shown in formula (8).

[0067] (8)

[0068] in, represents the hyperparameter of edge distance, s represents the set of positive triples in the collaborative knowledge graph, and accordingly, is a set of triplet negative samples, is a randomly sampled negative sample of t. Finally, the total loss function of this step can be expressed as .

[0069] S4.3 Intent-aware embedding generation: Embedding representation obtained from S4.1 and S4.2 and , through feature fusion functions (such as concatenation, summation or element-level attention mechanism), fusion and Generate intent-specific embeddings The fusion process can be expressed as formula (9).

[0070] (9)

[0071] in, is the fusion function. In this embodiment, the element-wise product method is used.

[0072] Typically, step S4.3 is followed by learning to align the embedding semantics through cross-network comparison.

[0073] Specifically, S4.4 dual network contrast learning aims to enhance the semantic consistency of the two network embeddings. , its intention embedding in the hypergraph and embedded in heterogeneous information networks Constitute a positive sample pair , and then randomly select its negative sample pair ,in Finally, the parameters of the InfoNCE loss optimization model are used to ensure that the embeddings generated by the two networks are semantically aligned. The parameters of the InfoNCE loss optimization model are

[0074]

[0075] in, is a similarity function (e.g., cosine similarity or vector dot product), is the temperature parameter, For items The cross-network negative sample set. Finally, the loss function of this part is .

[0076] S4.5 Perform intent decoupling classification based on the embedding of the specific intent.

[0077] To improve the discriminability of usage intent embeddings at a broader semantic level and reduce redundancy between different intent embeddings, this method introduces intent decoupling. Intents are classified using a classification model (such as a support vector machine, naive Bayes, or random forest), aiming to improve the model's overall ability to discriminate between different usage intents.

[0078] Specifically, this step is based on each node in S4.3 Intent-aware embedding To predict the usage intention category of the entire set. This process can be formalized as formula (11).

[0079] (11)

[0080] in, represents the predicted probability of each usage intention, and It is a multi-layer perceptron used to extract features from the disentangled embedding that are useful for intent classification.

[0081] In order to ensure the accuracy of the intent feature, this method uses the cross entropy loss function to optimize the model function, as shown in formula (12).

[0082] (12)

[0083] here, is an indicator function that takes the value of 1 when the true usage intention label matches the predicted label, and 0 otherwise. The design of this loss function can effectively improve the discriminative ability of the embedding vector of each usage intent in the semantic space.

[0084] Step 5: Aggregate the features of the user's historical purchases through the attention mechanism to obtain the user's embedding of intention k.

[0085] The purpose of this step is to generate a feature representation of the intention k for each user. , assuming that its historical purchase set is The present invention designs an attention network to aggregate each item in the purchase set. characteristics of the user, thereby obtaining About the embedding of intent k: .

[0086] Step 6: Calculate the intention relevance score between the user and the item, and generate a Top-N recommendation list by sorting the scores.

[0087] The purpose of this step is to generate recommendations for items that the user may be interested in. , the present invention uses a similarity function Calculate its comparison with all candidate items in the system The correlation score can be expressed as formula (13).

[0088] (13)

[0089] in, Candidate items Embedding y on intent k ij Comprehensively evaluate users from different perspectives With candidate items correlation. It can be cosine similarity, vector dot product, or Euclidean distance, etc. In this embodiment, vector dot product is used as an example. Finally, the relevance scores are sorted from high to low, and the top N items are selected as the recommendation list.

[0090] This part uses the cross entropy loss function to ensure the accuracy of the model's recommendation, as shown in formula (14).

[0091] (14)

[0092] Among them, when there is interaction between the user and the item, y ij The value is 1; otherwise, it is 0.

[0093] Preferably, this embodiment further includes step 7, constructing a joint optimization function to improve the performance of the entire method, and the loss values ​​obtained in S4.2, S4.4, S4.5 and S5 are weighted summed to obtain the total loss function L of the model, as shown in formula (15).

[0094] (15)

[0095] in, 、 and are all learnable parameters.

[0096] This paper develops an automated intent recognition method that can extract and quantify the performance characteristics of each item in different intent scenarios from user behavior data; based on the recognized intent information, a multi-dimensional intent feature vector is generated for each item to characterize its distribution characteristics in various intent dimensions; and precise matching recommendations are achieved by using intent vector similarity calculation, so that the system can accurately capture the user's current intent needs and generate personalized recommendation results accordingly.

[0097] At the same time, hypergraph and heterogeneous information network structures are constructed, including a hypergraph structure for modeling high-order associations between items, especially complex many-to-many interaction patterns, and a multi-level heterogeneous information network to effectively capture heterogeneous relationships between different types of entities and enhance the representation capability of the recommendation system. By combining the hypergraph and the heterogeneous information network, the hypergraph's modeling capability for high-order relationships is retained, while the heterogeneous information network's advantage in depicting interactions among multiple types of entities is fully utilized, thereby achieving all-round modeling of recommendation scenarios.

[0098] The present invention also includes an electronic device, including a memory, a processor and a computer program stored in the memory, wherein when the processor executes the program, the steps of any of the above-mentioned recommendation methods based on user intention perception are implemented.

[0099] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage and optical storage, etc.) containing computer-usable program code.

[0100] The above description is merely a preferred embodiment of the present application and is not intended to limit the present application. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present application shall be included within the scope of protection of the present application.

Claims

1. A recommendation method based on user intention perception, characterized in that: The following steps are involved: Step 1: Generate the initial embedding vector of the item: Map the ID of the item into the vector space to obtain the initial embedding vector of each item, which are recorded as , where d is the dimension of the embedding vector; Step 2: User behavior intention encoding, split the initial embedding vector into K intention sub-vectors, cluster the sub-vectors of all items under each intention, and generate the intention feature vector ; Step 3: Build a dual network, including building a hypergraph whose nodes are items and whose hyperedges include co-purchase hyperedges and intention hyperedges; build a heterogeneous information network whose nodes contain items and side information, and whose edges represent predefined relationships between nodes; Step 4: Intent decoding based on dual-graph representation learning, that is, adjusting the weight of the node in the hypergraph information propagation through the hypergraph attention network, so that the hyperedge with high semantic and intent matching with the current node is given a greater weight, including node-to-hyperedge attention aggregation and hyperedge-to-node attention aggregation, and updating the intent embedding of each node ; Update the nodes of the heterogeneous information network through the graph convolutional network, and use the triplet scoring mechanism to optimize and obtain the new node embedding Fusion and Generate an embedding of a specific intent; perform intent decoupling classification based on the embedding of the specific intent; Step 5: Aggregate the features of the user's historical purchases through the attention mechanism to obtain the user's embedding of intent k; Step 6: Calculate the intention relevance score between the user and the item, and generate a Top-N recommendation list by sorting the scores.

2. The recommendation method based on user intention perception according to claim 1, characterized in that: Cluster the sub-vectors of all items under each intent as described in step 2 to generate the intent feature vector Specifically, for the kth potential intent, the intent sub-vectors of all items are processed by clustering algorithm to generate several cluster centers, which are the intent feature vectors. .

3. The recommendation method based on user intention perception according to claim 1, characterized in that: In step 3, The process includes the following: filtering out intent hyperedges that are not related to intent k. When the similarity score between an item and intent k is greater than a predefined threshold, there is an intent hyperedge between the two, otherwise it does not exist.

4. The recommendation method based on user intention perception according to claim 1, characterized in that: The attention aggregation from the node to the hyperedge in step 4 includes the following: Aggregate all hyperedges j Connected node features, update the embedding representation of the hyperedge about intention k ,in, Indicates that under the k-th intention, the hyperedge The embedding vector of Represents the super edge midpoint The attention weight of The hyperedge-to-node attention aggregation includes, through the aggregation function Aggregate all hyperedge information associated with the current node to update the intent embedding of each node, ,in, It is a super edge For Node The attention weight of .

5. The recommendation method based on user intention perception according to claim 1, characterized in that: Fusion as described in step 4 and Generating intent-specific embeddings involves using a feature fusion function Fusion is performed to generate item embeddings with specific intent. The formula is: .

6. The recommendation method based on user intention perception according to claim 1, characterized in that: In the fusion and After generating the embedding for a specific intent, we also need to learn to align the embedding semantics through cross-network comparison. Specifically, for items , its intention embedding in the hypergraph and embedded in heterogeneous information networks Constitute a positive sample pair , and then randomly select its negative sample pair ,in Finally, the parameters of the InfoNCE loss optimization model are used to ensure that the embeddings generated by the two networks are semantically aligned. The parameters of the InfoNCE loss optimization model are ,in, is the similarity function, is the temperature parameter, For items The set of negative samples across the network.

7. The recommendation method based on user intention perception according to claim 1, characterized in that: The intent decoupling classification based on the embedding of the specific intent in step 4 includes classifying the intent using the classification model and classifying each node The intent embedding is updated.

8. The recommendation method based on user intention perception according to claim 1, characterized in that: The calculation of the user-item intention relevance score described in step 6 includes the similarity function Calculating users With all candidate items Relevance score ,in, For items Embedding on intent k, Represents a user Embedding on intent k.

9. The recommendation method based on user intention perception according to any one of claims 1 to 8, characterized in that: The method further includes step 7, constructing a joint optimization function to improve the performance of the method, wherein the loss function of the joint optimization function is the sum of the loss functions in step 4 and step 5.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory, wherein: When the processor executes the program, the steps of the recommendation method based on user usage intention perception as described in any one of claims 1 to 9 are implemented.

Citation Information

Patent Citations

  • Service modeling and recommendation method based on category perception graph neural network

    CN114817663A

  • Recommendation method based on deentanglement hypergraph contrast learning

    CN120086434A