A recommendation method based on double-intention modeling and intention guide knowledge graph denoising

Abstract

The present application relates to the technical field of intelligent recommendation and knowledge graph, and particularly relates to a recommendation method based on double-intent modeling and intent-guided knowledge graph denoising. The method constructs a user-item interaction graph and a knowledge graph, respectively learns group intent and semantic intent representations of users, and uses semantic intent to evaluate the importance of knowledge graph relationships, dynamically weights and suppresses noise in the knowledge propagation process. Then, the group intent representation and the semantic intent representation are fused to obtain the final representation of the user and the item, and preference prediction and recommendation are completed. The method solves the problem of insufficient intent modeling of existing knowledge graph recommendation models and the problem of knowledge graph noise propagation, improves the user interest modeling capability, recommendation accuracy and knowledge utilization efficiency, and can be widely applied to e-commerce, online video, music recommendation and other scenarios.

Description

Technical Field

[0001] This invention relates to the fields of recommendation systems and knowledge graph technology, specifically to a recommendation method based on dual-intent modeling and intent-guided knowledge graph denoising. Background Technology

[0002] With the rapid growth of internet information volume, recommender systems have become a key technology in e-commerce, online video platforms, social networks, and digital content services. Collaborative filtering methods achieve personalized recommendations by mining historical interaction relationships between users and items, but they generally suffer from problems such as data sparsity, cold start, and lack of interpretability. To enhance the semantic expressive power of recommender systems, researchers have introduced knowledge graphs into recommendation tasks, using information about entities and the relationships between them to assist in modeling user interests.

[0003] Existing knowledge graph recommendation methods typically learn user and item representations through graph neural networks or knowledge propagation mechanisms, and infer user preferences using entity relationships within the knowledge graph. However, most methods primarily explain user behavior at the semantic level, attributing user interactions to semantic information such as item attributes, categories, or entity relationships, while neglecting the collaborative patterns inherent in user group behavior. In reality, user decisions are often influenced by both long-term individual interests and group behavioral patterns; relying solely on semantic intent is insufficient to fully characterize users' true needs.

[0004] On the other hand, while knowledge graphs contain rich auxiliary information, they also contain a large number of entity relationships that are weakly or even irrelevant to the recommendation task. These noisy relationships can easily introduce invalid information during propagation, reducing the quality of user representations, increasing model computational overhead, and affecting the final recommendation performance. Existing knowledge graph denoising methods typically rely on complex graph pruning, graph augmentation, or contrastive learning mechanisms, which suffer from problems such as complex implementation, high training costs, and limited generalization ability.

[0005] Therefore, there is an urgent need for a recommendation method that can simultaneously model user semantic intent and group intent, and dynamically filter effective relationships in the knowledge graph for the recommendation task based on user interests, in order to improve the accuracy and robustness of the recommendation system. Summary of the Invention

[0006] The purpose of this invention is to provide a recommendation method based on dual-intent collaborative modeling and intent-guided knowledge graph denoising, which aims to solve the problems of existing knowledge graph recommendation models that only explain user behavior from a single semantic perspective, are difficult to fully explore the behavioral patterns of user groups, and are affected by the propagation of knowledge graph noise. This invention achieves multi-dimensional modeling of user interests and efficient utilization of knowledge graph information.

[0007] To achieve the above objectives, this invention proposes a recommendation method based on dual-intent collaborative modeling and intent-guided knowledge graph denoising. By simultaneously modeling user semantic intent and collective intent, and using semantic intent to dynamically guide the knowledge graph propagation process, it achieves accurate learning of user preferences and effective suppression of knowledge noise.

[0008] This invention constructs a heterogeneous information network composed of a user-item interaction graph and a knowledge graph. The user-item interaction graph describes the relationships between users' historical behaviors, while the knowledge graph describes the semantic relationships between item entities. By jointly utilizing the structural information of these two types of graphs, multi-perspective learning of user interest representations is achieved.

[0009] This invention designs a dual-intent collaborative modeling mechanism, including a semantic intent modeling module and a group intent modeling module. The semantic intent modeling module learns users' long-term stable interests and preferences. This module constructs a semantic propagation network on a knowledge graph, introduces multiple semantic intent prototype vectors, and learns the user's preference distribution across different semantic interest dimensions through the attention allocation relationship between user representations and semantic intent prototypes. It then uses this preference distribution to guide the aggregation of semantic information in the knowledge graph, thereby obtaining the user's semantic intent representation. The group intent modeling module learns the collaborative patterns inherent in user group behavior. This module constructs a group behavior propagation network on a user-item interaction graph, introduces multiple group intent prototypes, and learns the user's interest distribution under different group behavior patterns through the association relationship between users, items, and group intent prototypes. It then uses this interest distribution for neighborhood information propagation and aggregation, thereby obtaining the user and item group intent representations.

[0010] This invention posits that user decision-making behavior is typically influenced not only by long-term semantic interests but also by the combined effects of group behavioral patterns and collaborative preferences. Therefore, by simultaneously modeling semantic intent and group intent, user interests can be characterized from both individual preference and group behavior perspectives, achieving a more comprehensive and accurate user representation learning than single-interest modeling.

[0011] In the process of knowledge graph propagation, this invention further designs an intent-guided knowledge graph denoising mechanism. First, a mapping relationship between semantic intent and relation type is established, and an intent-relation weight matrix is ​​introduced to describe the importance of different semantic intents to various knowledge graph relations. Then, based on the user's current semantic intent distribution, a user-relation weight matrix is ​​calculated to obtain the contribution of various relations to user preferences.

[0012] During the knowledge dissemination phase, the user-relationship weight matrix is ​​used to dynamically weight the relationship propagation messages in the knowledge graph. For relationship types highly correlated with user interests, their propagation weight is increased; for relationship types less correlated with user interests, their propagation weight is decreased, thereby achieving noise suppression and effective knowledge preservation during the knowledge graph propagation process. Compared to traditional denoising methods such as graph pruning, graph augmentation, and contrastive learning, this invention achieves lightweight denoising of knowledge graphs without introducing additional complex network structures, improving model efficiency and scalability.

[0013] To achieve the above objectives, this invention proposes a recommendation method based on dual-intent collaborative modeling and intent-guided knowledge graph denoising, comprising the following steps:

[0014] Step 1: Obtain user historical interaction data, construct a user-item interaction graph, and obtain knowledge graph data to construct an entity relationship graph;

[0015] Step 2: Obtain knowledge graph data, construct entity relationship graph, and define a dual-intent recommendation prediction task under the joint constraints of user-item interaction graph and knowledge graph;

[0016] Step 3: Initialize the embedded representations of users, items, entities, and relationships;

[0017] Step 4: Construct a set of semantic intent prototypes and propagate the learned user semantic intent representations on the interaction graph;

[0018] Step 5: Construct a set of group intent prototypes and propagate the learned user group intent representations and item group intent representations on the user-item interaction graph;

[0019] Step 6: Establish an intent-relationship mapping matrix and calculate the importance of users to different knowledge graph relationship types;

[0020] Step 7: Dynamically weight the messages propagated in the knowledge graph based on the importance of relationships to achieve denoising of the knowledge graph guided by intent;

[0021] Step 8: Merge semantic intent representation and group intent representation to obtain end-user representation and end-item representation;

[0022] Step 9: Calculate the user preference prediction score based on the end-user representation and the final item representation;

[0023] Step 10: Jointly train the model using the ranking optimization objective and intent relevance constraints;

[0024] Step 11: Generate a user preference ranking list using the trained model, sorting them from highest to lowest based on predicted scores, and outputting Top-K recommendation results. Compared with existing technologies, this invention can simultaneously characterize users' semantic preferences and group behavior patterns, and suppress the propagation of weakly related relationships during knowledge graph propagation, thereby improving the accuracy and stability of recommendation results. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 The flowchart of a dual-intent recommendation method according to the present invention is shown in the abstract figure.

[0027] Figure 2 This is a schematic diagram illustrating the denoising methodology intended to guide this invention. Detailed Implementation

[0028] The recommendation method proposed in this invention, based on dual-intent collaborative modeling and intent-guided knowledge graph denoising, includes the following steps:

[0029] Step 1: Obtain user historical interaction data, construct a user-item interaction graph, and obtain knowledge graph data to construct an entity relationship graph. Let the user set be... The collection of items is The user-item interaction edge set is Then the user-item interaction graph is represented as Among them, if the user With items If there are historical interactions, then establish edges in the interaction graph. The user-item interaction matrix is ​​defined as follows:

[0030]

[0031] Step 2: Obtain knowledge graph data and construct an entity relationship graph. The knowledge graph is represented as follows:

[0032]

[0033] in, Indicates the head entity. Indicates the tail entity. Indicates the relation type, Represents a set of entities. Represents a set of relation types.

[0034] The dual-intention recommendation task is defined as a recommendation prediction function constrained by both the user-item interaction graph and the knowledge graph:

[0035]

[0036] in, Indicates user, Indicates candidate items, Represents the set of model parameters. Indicates user For candidate items The predicted preference score.

[0037] Step 3: Initialize the embedded representations of users, items, entities, and relationships.

[0038] Let the user embedding matrix be The item embedding matrix is The entity embedding matrix is The relation embedding matrix is ,in This represents the dimension of the embedding vector.

[0039] Step 4: Construct a set of semantic intent prototypes and propagate the learned user semantic intent representations on the interaction graph.

[0040] Specifically, introducing A semantic intent prototype matrix is ​​used to represent a user's long-term, stable semantic interest preferences.

[0041]

[0042] in, Indicates the first A semantic intent prototype vector Indicates the number of semantic intent prototypes. Indicates the dimension of the embedding vector.

[0043] In the During layer propagation, the current semantic representations of all users are combined into a user semantic representation matrix. And based on the correlation between the user's semantic representation and the semantic intent prototype, the user semantic intent attention matrix is ​​calculated:

[0044]

[0045] in, Indicates the first The semantic representation matrix of all users in the layer; Represents the semantic intent prototype matrix; express The transpose of the matrix; express Normalization function; Indicates the first The semantic intent attention matrix for all users in the layer. For each user... , Indicates user In the The weight distribution of each semantic intent prototype corresponds to the layer.

[0046] Furthermore, based on users The semantic intent attention weights are calculated by weighting and summing the semantic intent prototypes to obtain the user's... Intent-aware semantic representation:

[0047]

[0048] in, Indicates user In the Layer-based intention-aware semantic representation Indicates user For the Attention weights for each semantic intent prototype Indicates the first A semantic intent prototype vector.

[0049] Subsequently, the semantic aggregation process of user-interacted items is modulated using the intent-aware semantic representation to obtain the next layer of user semantic representation:

[0050]

[0051] in, Indicates user In the Semantic representation of layers; Indicates user The set of neighboring items, i.e., the user's A collection of items that have been interacted with; This indicates the number of items in the neighbor's item set; Represents items In the Semantic representation of layers; This indicates element-wise multiplication.

[0052] Through the above processing, the user semantic intent representation is not only used to characterize the user's preference distribution in the semantic space of the knowledge graph, but also to guide the semantic aggregation on the user side, so that the user representation can highlight the semantic information of items that are more consistent with their semantic interests.

[0053] Step 5: Construct a set of group intent prototypes and propagate the learned user group intent representations and item group intent representations on the user-item interaction graph.

[0054] Specifically, in order to learn the group behavior patterns in the user-item interaction graph, this invention introduces... A group intent prototype. Unlike the semantic intent module, which utilizes structured relationship information in a knowledge graph, the group intent module directly draws from the user-item interaction graph. Learning collaborative behavior patterns.

[0055] Let the first The user group representation matrix of the layer is as follows , No. The item group representation matrix of the layer is To characterize the different collaborative roles of the user and the item in the interaction graph, we introduce a user-side group intent prototype matrix and an item-side group intent prototype matrix:

[0056]

[0057]

[0058] in, This represents the prototype matrix of user-side group intent. Indicates the first A prototype vector of user-side group intent; This represents the prototype matrix of the group's intentions on the object side. Indicates the first A prototype vector of the group's intention on the item side; This indicates the number of prototypes of group intentions.

[0059] In the During the layered propagation process, the user-side group intent attention matrix is ​​first calculated based on the correlation between the user group representation and the user-side group intent prototype:

[0060]

[0061] in, Represents the prototype matrix of user-side group intent The transpose of the matrix; Represents the normalization function; Indicates the first The attention weight matrix of all users on the user-side group intent prototype. For users , Indicates user Weight distribution of intent prototypes for each user group.

[0062] Furthermore, based on users The group intent attention weight is calculated by weighting and summing the user-side group intent prototypes to obtain the user's attention weight. Intention perception group representation:

[0063]

[0064] in, Indicates user In the Layered intention perception group representation, Indicates user For the Attention weights of user-side group intent prototypes Indicates the first A prototype of user-side group intent.

[0065] Subsequently, the intent-aware group representation is used to modulate the group information aggregation process of items that the user has interacted with, resulting in the next layer of user group representation:

[0066]

[0067] in, Indicates user In the Group representation of layers; Indicates user The set of neighboring items, i.e., the user's A collection of items that have been interacted with; This indicates the number of items in the neighbor's item set; Represents items In the Group representation of layers; This indicates element-wise multiplication.

[0068] Corresponding to the user-side propagation process, on the item side, based on the correlation between the item group representation and the item-side group intent prototype, the item-side group intent attention matrix is ​​calculated:

[0069]

[0070]

[0071] in, Represents the prototype matrix of group intentions on the item side. The transpose of the matrix; Indicates the first The attention weight matrix for all items in the layer on the item-side group intent prototype. For each item... , Represents items Weight distribution of the group intention prototype for each item.

[0072] Furthermore, based on the items The group intention attention weights are calculated by weighting and summing the group intention prototypes on the item side to obtain the item's... Intention perception group representation:

[0073]

[0074] in, Represents items In the Layered intention perception group representation, Represents items For the Attention weights of object-side group intention prototypes Indicates the first A prototype of the intention of a group of objects.

[0075] Subsequently, the intent-aware group representation is used to modulate the aggregation process of user information interacting with item i, resulting in the next layer of item group representation:

[0076]

[0077] in Represents items In the Group representation of layers; Represents items The set of neighboring users, that is, the set of users related to the item The set of users who interacted; This indicates the number of users in the neighbor user set; Indicates user In the Group representation of layers.

[0078] In this way, the present invention maintains the user-side group intent prototype and the item-side group intent prototype respectively, so that the user-item interaction is modeled as a matching relationship between user-side preferences and item-side collaborative roles, thereby obtaining a more expressive collaborative recommendation representation.

[0079] Step 6: Establish an intent-relationship mapping matrix and calculate the importance of users to different knowledge graph relationship types.

[0080] Specifically, in order to characterize the degree of association between different semantic intentions and knowledge graph relation types, this invention constructs an intention-relationship weight matrix:

[0081]

[0082] in, Intended meaning—relationship weight matrix; Indicates the first A semantic intent for relation types The degree of importance; Indicates the number of semantic intent prototypes; Represents the set of relation types in a knowledge graph; This indicates the number of relation types. In a specific implementation, the intent-relationship weight matrix... From the trainable parameter matrix go through Normalization yields:

[0083]

[0084] in, Represents a trainable intent-relationship scoring matrix. This represents the normalization function performed along the relation type dimension.

[0085] Furthermore, in conjunction with the first Layered user semantic intent attention matrix Intent-Relationship Weight Matrix Calculate the user-relationship weight matrix:

[0086]

[0087] in, Indicates the first Layered user semantic intent attention matrix; Indicates the first Layered user-relationship weight matrix; Indicates user In the Hierarchical Relationship Types The importance assessment results are used to map the distribution of user-side semantic intent to the user's preference weights for various relationships in the knowledge graph.

[0088] Step 7: Dynamically weight the messages propagated in the knowledge graph by utilizing the importance of relationships to achieve denoising of the knowledge graph guided by intent.

[0089] Specifically, in order to obtain global-level relation denoising weights, this invention modifies the user-relationship weight matrix. The average of the relationship weights of all users is used to obtain the first... Global relation importance vector of the layer:

[0090]

[0091] in, Represents the user-relationship weight matrix The Line, i.e., user Distribution of importance among different relation types; Indicates the first Layer-wide relation importance vector. Relationships with lower importance have their weight reduced during subsequent knowledge graph propagation, while relationships with higher importance have their weight increased during propagation.

[0092] For knowledge graphs any triplet in This invention is based on relation type Global importance is used to construct relation-aware semantic message propagation:

[0093]

[0094] in, Representing relation type In the The global importance of the layer Indicates the first Layer consists of tail entities Through relationships propagation to head entity Relationship-aware semantic messages; Represents tail entity In the Semantic embedding representation of layers; Representing relation type Relational embedding vector; This indicates element-wise multiplication.

[0095] Furthermore, for entities in the knowledge graph Based on its set of neighbor triples Average aggregation of relation-aware semantic messages yields entities. Semantic representation at the next level:

[0096]

[0097] in, Representing entities In the Semantic representation of layers; Representing entities The set of neighbor triples in a knowledge graph; This indicates the number of triples in the set of neighbor triples.

[0098] Since items in the recommendation system are also included as entities in the knowledge graph entity set. Therefore, for items Its next-level semantic representation can be updated in the following way:

[0099]

[0100] in, Represents items In the Semantic representation of layers; Represents items The set of neighbor triples in a knowledge graph; This indicates the number of triples in the set of neighbor triples.

[0101] Step 8: Merge semantic intent representation and group intent representation to obtain end-user representation and end-item representation.

[0102] Specifically, after completing the semantic intent propagation, group intent propagation, and intent-guided knowledge graph denoising, the user's group representation and semantic representation are concatenated to obtain the user's... The final representation:

[0103]

[0104] in, Indicates user The final representation; Indicates user The expression of group intent; Indicates user The semantic intent representation; This indicates a vector concatenation operation.

[0105] Accordingly, the group representation of the item is concatenated with its semantic representation to obtain the item. The final representation:

[0106]

[0107] in, Represents items The final representation; Represents items The expression of group intent; Represents items The semantic intent is represented.

[0108] Therefore, the final user representation and the final item representation simultaneously contain collaborative preference information from the user-item interaction graph and semantic association information from the knowledge graph, thereby improving the expressive power of the recommendation representation.

[0109] Step 9: Calculate the user preference prediction score based on the end-user representation and the end-item representation.

[0110] Specifically, in acquiring users The final expression and candidate items The final expression Then, the user is calculated using the inner product. For candidate items Preference matching scores, and utilize The activation function yields the predicted probability of an interaction between the user and an item:

[0111]

[0112] in, Indicates user With items The predicted probability or preference prediction score of the interaction between them; This indicates the user's final expression. transpose; Represents items The final representation; express Activation function.

[0113] During the recommendation phase, for users It can calculate the preference prediction scores between itself and multiple candidate items. The candidate items are then sorted from highest to lowest according to their preference prediction scores to obtain a user-oriented result. The recommended results.

[0114] Step 10: Jointly train the model using the ranking optimization objective and intent relevance constraints, and output the recommendation results. The Bayesian personalized ranking loss function is used to optimize the user's preference ranking relationship for positive and negative feedback items. The calculation formula is as follows:

[0115]

[0116] in, Represents the set of training sample triples; Indicates user; Indicates user Positive feedback items; Indicates user Negative feedback items; Indicates user For positive feedback items Predicted score; Indicates user For negative feedback items Predicted score; Represents the set of model parameters; Represents the regularization coefficient; Representing model parameters Regular terms.

[0117] Furthermore, to reduce information redundancy between different intent representations, this invention introduces an intent relevance constraint term, the calculation result of which is denoted as:

[0118]

[0119] in, Represents the weighting coefficients of the intent relevance constraint terms; Represents the semantic intent prototype matrix; This represents the prototype matrix of user-side group intent. Represents the prototype matrix of the group's intention on the object side; A measure of the relevance between semantic intent prototypes; This represents a measure of the correlation between user-side group intent prototypes; This represents a measure of the correlation between prototypes of group intentions on the object side. Therefore, the overall training objective of this invention is:

[0120]

[0121] Step 11: Use the trained model to generate a user preference ranking list, sort it from high to low according to the predicted score, and output the Top-K recommendation results.

[0122] In summary, this invention achieves unified processing of long-term user semantic preferences, group collaborative behavior patterns, and knowledge graph relation noise through the synergistic effect of semantic intent modeling, group intent modeling, and intent-guided knowledge graph denoising. Compared with recommendation methods that rely solely on knowledge graph semantic propagation or single collaborative filtering, this invention can more comprehensively characterize user decision-making motivations and suppress the propagation of irrelevant relations with lower additional complexity, thereby improving the accuracy and stability of recommendation results.

[0123] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A recommendation method based on dual-intent collaborative modeling and intent-guided knowledge graph denoising, characterized in that, Performed by a computing device, the process includes the following steps: S1, acquire user historical interaction data, construct user-item interaction graph, and acquire knowledge graph data to construct entity relationship graph, wherein the user-item interaction graph is used to represent the historical interaction relationship between users and items, and the entity relationship graph is used to represent entities, relationship types and triple relationships between entities, and items are mapped to entities in the entity relationship graph; S2, Based on the user-item interaction graph and the entity relationship graph, define a dual-intent recommendation prediction task under the common constraints of the two; S3 initializes the embedded representations of users, items, entities, and relationships; S4, construct a set of semantic intent prototypes, calculate user semantic intent attention based on the correlation between user semantic representation and semantic intent prototypes, obtain user semantic intent representation based on the user semantic intent attention, and aggregate the semantic representation of user-interacted items by weighting each element using the user semantic intent representation. S5, construct a set of user-side group intention prototypes and a set of item-side group intention prototypes, and propagate the learned user group intention representation and item group intention representation on the user-item interaction graph; S6, establish an intent-relationship mapping matrix, and calculate the importance of the user to different knowledge graph relationship types by combining the user semantic intent attention; S7, dynamically weight the knowledge graph propagation message according to the importance of the relationship type to achieve denoising of the knowledge graph guided by intent, and update the semantic representation of entities and the semantic representation of items; S8 integrates semantic intent representation and group intent representation to obtain end-user representation and end-item representation; S9, Calculate the user's preference prediction score for candidate items based on the end-user representation and the end-item representation; S10, the model is jointly trained using the ranking optimization objective and the intent relevance constraint; S11. Use the trained model to generate a user preference ranking list, sort the preferences from high to low according to the predicted scores, and output the recommendation results.

2. The method according to claim 1, characterized in that, In step S4, the semantic intent prototype matrix includes multiple semantic intent prototype vectors. The user semantic intent attention is obtained by normalizing the product between the l-th layer user semantic representation matrix and the transpose of the semantic intent prototype matrix. The user semantic intent representation is the weighted sum of each semantic intent prototype according to the user semantic intent attention, and the semantic embedding of the user-interacted items is aggregated element-wise using the user semantic intent representation to obtain the next layer user semantic representation.

3. The method according to claim 1, characterized in that, In step S5, user-side group intent attention is calculated based on the correlation between user group representation and user-side group intent prototype, and item-side group intent attention is calculated based on the correlation between item group representation and item-side group intent prototype. User intent-aware group representation is obtained based on the user-side group intent attention, and the group representations of adjacent items are aggregated after element-wise weighting using the user intent-aware group representation. Item intent-aware group representation is obtained based on the item-side group intent attention, and the group representations of adjacent users are aggregated after element-wise weighting using the item intent-aware group representation.

4. The method according to claim 1, characterized in that, In step S6, the intent-relationship mapping matrix is ​​obtained by normalizing the trainable intent-relationship scoring matrix, and the importance of the user to different knowledge graph relationship types is obtained by multiplying the user semantic intent attention matrix with the intent-relationship mapping matrix. In step S7, the relationship weights of all users are averaged to obtain a global relationship importance vector, and for triples in the entity relationship graph, the element-wise combination results of tail entity semantic embedding and relationship embedding are weighted according to the global relationship importance weight corresponding to the relationship type of the triple to obtain a relationship-aware semantic propagation message. Step S7 also includes: aggregating the relationship-aware semantic propagation messages in the neighbor triple set of the entity to obtain the next-layer entity semantic representation; when the entity is an item, the relationship-aware semantic propagation messages in the neighbor triple set of the item are aggregated to obtain the next-layer item semantic representation.

5. The method according to claim 1, characterized in that, In S8 and S9, the user group intent representation and the user semantic intent representation are concatenated to obtain the final user representation, and the item group intent representation and the item semantic intent representation are concatenated to obtain the final item representation. The user's preference prediction score for the candidate items is calculated by the inner product between the final user representation and the final item representation and the activation function.

6. The method according to claim 1, characterized in that, In S10, the ranking optimization objective includes Bayesian personalized ranking loss, and the intent relevance constraint includes relevance regularization terms calculated on semantic intent prototypes, user-side group intent prototypes, and item-side group intent prototypes, and the weighted sum of the Bayesian personalized ranking loss and the relevance regularization terms is used as the overall training objective.

7. A recommendation device based on dual-intent collaborative modeling and intent-guided knowledge graph denoising, characterized in that, include: The data construction module is used to acquire user historical interaction data and knowledge graph data to construct user-item interaction graphs and entity relationship graphs; The task definition and initialization module is used to define the dual-intent recommendation and prediction task and initialize the embedded representations of users, items, entities, and relationships. The semantic intent module is used to perform the processing in S4 of claim 1; the group intent module is used to perform the processing in S5 of claim 1; and the intent-guided denoising module is used to perform the processing in S6 and S7 of claim 1. A fusion prediction module is used to perform the processing in S8 and S9 of claim 1; The training module is used to perform the processing in S10 of claim 1; The recommended output module is used to perform the processing in S11 of claim 1.

8. A computing device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, implements the method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, It stores a computer program thereon, which, when executed by a processor, implements the method described in any one of claims 1 to 6.

Images