CN-121980518-A - Recommendation method, device, computer equipment and storage medium based on graph migration

CN121980518ACN 121980518 ACN121980518 ACN 121980518ACN-121980518-A

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a recommendation method, a device, computer equipment and a storage medium based on graph migration, wherein the method comprises the steps of obtaining original business data, constructing an abnormal graph based on the original business data, obtaining auxiliary features, mapping initial features and auxiliary features of nodes in the abnormal graph to obtain graph node embedding and auxiliary feature embedding, injecting semantic information in the auxiliary feature embedding into the graph node embedding to obtain migration enhancement graph embedding, performing graph volume aggregation on the migration enhancement graph embedding to obtain enhancement graph features, fusing the enhancement graph features and the auxiliary feature embedding to obtain target fusion features, predicting the preference of a user on an object based on the target fusion features to obtain preference information, and recommending the object based on the preference information. The method effectively relieves the data sparsity problem and improves the accuracy and the robustness of the recommendation system in a cold start scene.

Inventors

WANG JIANZONG
DENG YUWEI

Assignees

平安科技（深圳）有限公司

Dates

Publication Date: 20260505
Application Date: 20260302

Claims (10)

1. A graph migration-based recommendation method, comprising: Acquiring original business data, constructing an abnormal composition containing user nodes, object nodes and attribute nodes based on the original business data, and acquiring auxiliary features associated with the user nodes and/or the object nodes; Mapping the initial features and the auxiliary features of the nodes in the different patterns to feature spaces with the same dimension respectively to obtain pattern node embedding and auxiliary feature embedding; injecting the semantic information in the auxiliary feature embedding into the graph node embedding through an attention migration mechanism to obtain a migration enhancement graph embedding; performing graph convolution aggregation on the migration enhancement graph embedding based on the topological structure of the different graph to obtain enhancement graph characteristics; Fusing the reinforcement map features with the auxiliary feature embedding to obtain target fusion features; predicting the preference of the user for the articles based on the target fusion characteristics, obtaining preference information, and recommending the articles to the user based on the preference information.
2. The graph migration-based recommendation method of claim 1, wherein the obtaining the original business data and constructing an iso-graph comprising user nodes, item nodes, and attribute nodes based on the original business data, and obtaining auxiliary features associated with the user nodes and/or the item nodes, comprises: Acquiring the original business data, and constructing user nodes, article nodes and attribute nodes based on each user, article and attribute in the original business data; According to the historical interaction data of the user and the article in the original business data, constructing an interaction edge between the user node and the article node; constructing a home edge between the article node and the attribute node according to the attribute information of the article in the original service data; Constructing an associated edge between the user node and the attribute node according to the portrait information of the user in the original service data; Taking the attribute node as an intermediary, and constructing a second-order association path for connecting the user node and the object node which are not directly interacted with each other; and constructing the heterogeneous graph based on the user node, the object node, the attribute node, the home edge, the association edge and the second-order association path, and acquiring auxiliary features associated with the user node and/or the object node.
3. The graph migration-based recommendation method according to claim 1, wherein mapping the initial feature and the auxiliary feature of the node in the abnormal graph to feature spaces of the same dimension respectively to obtain graph node embedding and auxiliary feature embedding comprises: Mapping initial characteristics of nodes in the iso-graph to a characteristic space with preset dimensions through a graph initialization embedding function to generate low-dimensional dense graph node embedding; encoding the auxiliary features according to categories to generate auxiliary feature vectors; And mapping the auxiliary feature vector to the feature space of the preset dimension to generate the auxiliary feature embedding.
4. The graph migration-based recommendation method according to claim 1, wherein the injecting semantic information in the auxiliary feature embedding into the graph node embedding through an attention migration mechanism to obtain a migration-enhanced graph embedding includes: calculating an attention weight matrix between the graph node embedding and the auxiliary feature embedding; carrying out weighted summation on the auxiliary feature embedding according to the attention weight matrix to obtain a weighted summation result; and adding the weighted summation result and the graph node embedding to obtain the migration enhanced graph embedding.
5. The graph migration-based recommendation method of claim 4, wherein when a node in the iso-graph is identified as a sparse node, the computing an attention weight matrix between the graph node embedding and the assist feature embedding comprises: The semantic similarity between the sparse nodes and the auxiliary features embedded in the corresponding samples is used as a bias term; the attention weight matrix between the graph node embedding and the assist feature embedding is calculated based on the bias term.
6. The graph migration based recommendation method according to any one of claims 1 to 5, wherein when a node in the iso-graph is not identified as a sparse node, the fusing the reinforcement graph feature with the auxiliary feature embedding to obtain a target fusion feature includes: Multiplying the reinforcement map features and the auxiliary feature embedding element by element to obtain interaction features; and embedding and adding the interaction feature, the enhancement map feature and the auxiliary feature to obtain the target fusion feature.
7. The graph migration-based recommendation method according to any one of claims 1 to 5, wherein when a node in the iso-graph is identified as a sparse node, the fusing the enhanced graph feature with the assist feature embedding to obtain a target fusion feature includes: Dynamically calculating the fusion weight of the reinforcement graph features and the auxiliary feature embedding according to the sparsity of the sparse nodes; And carrying out weighted fusion on the enhancement map features and the auxiliary feature embedding according to the fusion weights to generate the target fusion features.
8. A graph migration-based recommendation device, comprising: the system comprises an abnormal composition construction module, a data processing module and a data processing module, wherein the abnormal composition construction module is used for acquiring original service data, constructing an abnormal composition comprising user nodes, object nodes and attribute nodes based on the original service data, and acquiring auxiliary characteristics associated with the user nodes and/or the object nodes; the feature embedding module is used for mapping the initial features and the auxiliary features of the nodes in the abnormal composition to feature spaces with the same dimension respectively to obtain graph node embedding and auxiliary feature embedding; the feature migration module is used for injecting semantic information in the auxiliary feature embedding into the graph node embedding through an attention migration mechanism to obtain a migration enhancement graph embedding; the graph convolution aggregation module is used for performing graph convolution aggregation on the migration enhancement graph embedding based on the topological structure of the different graph to obtain enhancement graph characteristics; The feature fusion module is used for embedding and fusing the enhancement map features and the auxiliary features to obtain target fusion features; And the item recommending module is used for predicting the preference of the user on the item based on the target fusion characteristic, obtaining preference information and recommending the item to the user based on the preference information.
9. A computer device comprising a memory and a processor, the memory having stored therein a computer program, the processor implementing the graph migration based recommendation method according to any one of claims 1 to 7 when the computer program is executed.
10. A computer readable storage medium, wherein a computer program is stored on the computer readable storage medium, which when executed by a processor implements the graph migration based recommendation method according to any one of claims 1 to 7.

Description

Recommendation method, device, computer equipment and storage medium based on graph migration Technical Field The application relates to the technical field of artificial intelligence, and can be applied to the medical field and the financial science and technology field, in particular to a recommendation method, a recommendation device, a computer device and a storage medium based on graph migration. Background The application of personalized recommendation technology in key fields such as financial science and technology and intelligent medical treatment is continuously deepened, however, the existing method exposes systematic defects when dealing with the core challenges of the fields. In a financial science and technology scenario, intelligent consultation and credit recommendation services frequently encounter cold start dilemma of new users and products, and when users lack historical behavioral data or the products have not accumulated enough interaction records, it is difficult for a recommendation system to build an effective user preference model. Meanwhile, financial business relates to deep fusion requirements of multi-dimensional heterogeneous features, including user portrait information, dynamic market data and product attribute features. The features have obvious semantic difference and structural heterogeneity, traditional collaborative filtering and other methods excessively depend on dense user-product interaction data, the performance is rapidly deteriorated under the data sparseness condition, and complex nonlinear association among the features cannot be sufficiently captured, so that the recommendation result lacks accuracy and personalized adaptation capability. In the smart medical field, auxiliary diagnosis and treatment recommendation face more serious challenges. Patient sample size is extremely deficient in rare disease diagnosis and treatment scenes, and medical information itself forms a highly complex multi-modal knowledge network, covering structured electronic medical records, unstructured medical image data, genomic information and drug interaction maps. The prior art is difficult to effectively integrate symptom descriptions in medical record texts, pathological features in image data and biomarkers in gene sequences, and cannot model deep medical logic association, such as causal relationship between disease evolution paths and treatment responses. Although the neural network-based recommendation model can partially utilize the structural information of the user-item interaction graph, it has fundamental limitations in practical deployment. On the one hand, the fusion capability of the models to the multi-source heterogeneous features is insufficient, and only shallow feature stitching or simple weighting is usually carried out, so that deep semantic complementation between the structural features and auxiliary features of the graph cannot be realized, and information loss of feature expression is caused. On the other hand, the existing method lacks an adaptive processing mechanism for data sparse nodes, and when user nodes or object nodes in the graph are in a sparse state due to sparse interactive data, the feature learning strategy cannot be dynamically adjusted, so that the model shows poor practicability and robustness in real scenes such as financial new customer recommendation or rare disease diagnosis and treatment. Therefore, development of a novel recommendation framework is needed to deeply integrate multi-source heterogeneous characteristics, intelligently relieve the influence of data sparsity, and optimize the calculation efficiency while guaranteeing the recommendation precision so as to support the actual business demands in key fields such as financial science and technology, intelligent medical treatment and the like. Disclosure of Invention The embodiment of the application aims to provide a recommendation method, a recommendation device, computer equipment and a storage medium based on graph migration, which effectively relieve the problem of data sparseness and improve the accuracy and the robustness of a recommendation system in a cold start scene. In order to solve the above technical problems, an embodiment of the present application provides a recommendation method based on graph migration, including: Acquiring original business data, constructing an abnormal composition containing user nodes, object nodes and attribute nodes based on the original business data, and acquiring auxiliary features associated with the user nodes and/or the object nodes; Mapping the initial features and the auxiliary features of the nodes in the different patterns to feature spaces with the same dimension respectively to obtain pattern node embedding and auxiliary feature embedding; injecting the semantic information in the auxiliary feature embedding into the graph node embedding through an attention migration mechanism to obtain a migration enhancement graph embeddi