CN-121562737-B - Model training method, recommendation method and system for cross-border recommendation graph neural network

CN121562737BCN 121562737 BCN121562737 BCN 121562737BCN-121562737-B

Abstract

The application relates to the technical field of graph neural networks, in particular to a model training method, a recommendation method and a system for recommending a graph neural network in a cross-border mode. The cross-border recommendation model trained by the model training method is used for generating characteristic representations of the target object and the target item, the cross-border recommendation model aggregates first embedded vectors of second-order item neighbors of the target object, the second-order item neighbors are connected with the target object through implicit association objects, and aggregates second embedded vectors of second-order object neighbors of the target item, and the second-order object neighbors are connected with the target item through implicit association items. According to the method, new implicit relation aggregation constraint is put forward, the capability of learning generalizable characterization from sparse and isolated local data under a non-overlapping cross-border scene of an object is enhanced, so that each client forms a cross-border recommendation model with highly overlapped parameter spatial distribution, and parameters of a global model aggregated at a server can be converged.

Inventors

TAN ZHIZHONG
ZHANG FUBO
GONG DEQUAN
WAN XUANCHEN
CHEN TINGYI
Wang gulin

Assignees

广东省智能科学与技术研究院

Dates

Publication Date: 20260505
Application Date: 20260120

Claims (8)

1. A model training method of a cross-border recommendation graph neural network, which is characterized by being applied to a client of federal learning, the method comprising: determining an object interaction sequence based on the historical interaction record of the target object, and determining an item interaction sequence based on the historical interaction record of the target item; Constructing a dynamic bipartite graph through the object interaction sequence and the project interaction sequence, wherein the dynamic bipartite graph comprises bipartite graphs corresponding to N time steps, and N is an integer greater than or equal to 1; Constructing or updating a cross-border recommendation model, wherein in the process of transmitting information on the dynamic bipartite graph, the cross-border recommendation model aggregates first embedded vectors of second-order item neighbors of the target object, the second-order item neighbors are connected with the target object through implicit association objects, and aggregates second embedded vectors of second-order object neighbors of the target object, and the second-order object neighbors are connected with the target object through implicit association objects; Training the cross-border recommendation model based on the training triples and the dynamic bipartite graphs, and sending model update data to a server so that the server can determine a global model according to the cross-border recommendation models of a plurality of clients; Building a dynamic bipartite graph through the object interaction sequence and the item interaction sequence, wherein the dynamic bipartite graph comprises the following specific steps: converting each interaction behavior in the object interaction sequence into an object dense vector, and embedding the object dense vector according to time attenuation to obtain an embedded representation of the target object; Converting each interaction behavior in the item interaction sequence into an item dense vector, and embedding the item dense vector according to time attenuation to obtain an embedded representation of the target item; Constructing the dynamic bipartite graph according to the embedded representation of the target object and the embedded representation of the target item; the cross-border recommendation model is provided with a GNN module and a perception module; the training of the cross-border recommendation model based on the training triples and the dynamic bipartite graph specifically comprises the following steps: Encoding the training triples through interaction scoring tags, and determining an adjacency matrix through the dynamic bipartite graph; The method comprises the steps of inputting the training triples and the adjacency matrix after coding into the cross-border recommendation model, wherein the GNN module aggregates first embedded vectors of second-order item neighbors of the target object along a first meta-path to obtain target object feature representation; and updating the cross-border recommendation model according to the interaction score label and the prediction loss determined by the interaction score.
2. The model training method of claim 1, wherein the GNN module aggregates the first embedded vectors of the second order item neighbors of the target object along a first meta-path to obtain a target object feature representation, comprising: determining a common following object according to the target object and the first-order object neighbor of the target object; Determining a first-order object neighbor of the common following object as an implicit associated object of the target object, and transmitting an embedded vector of a first-order item neighbor of the implicit associated object as a first embedded vector; And aggregating all the first embedded vectors to obtain the target object characteristic representation.
3. The model training method of claim 1, wherein the GNN module aggregates second embedded vectors of second order object neighbors of the target item along a second binary path to obtain a target item feature representation, comprising: determining a first-order item neighbor of the target item according to a common scoring object set, wherein the first-order item neighbor is an item with interactive records with objects in the common scoring object set; Determining a first-order item neighbor of the target item as the implicit association item, and transmitting an embedded vector of the first-order object neighbor of the implicit association item as a second embedded vector; and aggregating all the second embedded vectors to obtain the target item feature representation.
4. A model training method as claimed in claim 3, wherein determining first order item neighbors of the target item from a set of common scoring objects comprises: determining a common scoring object set through historical interaction objects of the target item and the candidate item; calculating interaction strength of each object in the common scoring object set on the target item and the candidate item; And determining whether the candidate item is determined to be a first-order item neighbor of the target item according to the interaction strength.
5. The model training method of claim 4, wherein calculating the interaction strength of each object in the common scoring object set with respect to the target item and the candidate item comprises: Determining scale confidence according to the number of objects of the common scoring object set; acquiring a first score of each object in the common scoring objects on the target item and a second score of each object on the candidate item, and calculating the distance between the first score and the second score to determine evaluation similarity; And determining the interaction strength according to the scale confidence and the evaluation similarity.
6. A method for training a model of a neural network of a cross-border recommendation graph, which is applied to a server of federal learning, and the cross-border recommendation model obtained by the method of any one of claims 1 to 5 is aggregated at the server, and the method comprises: obtaining model update data of the cross-border recommendation model uploaded by K clients, wherein K is an integer greater than or equal to 1; and determining a global model according to the model updating data of the cross-border recommendation model.
7. A cross-border object recommending method is characterized in that, the method for recommending the target item to the target object based on the global model obtained by the model training method of claim 6 specifically comprises the following steps: inputting an interaction triplet into the global model to obtain an interaction score, wherein the interaction triplet comprises the target object, the target item and a time step, and the interaction score characterizes the interest degree of the target object to the target item; And if the interaction score is larger than a preset interaction score threshold, pushing the target item to the target object.
8. A cross-border object recommendation system comprising a processor and a memory having stored therein computer readable execution code which when executed by the processor implements the method of any of claims 1 to 7.

Description

Model training method, recommendation method and system for cross-border recommendation graph neural network Technical Field The application relates to the technical field of graph neural networks, in particular to a model training method, a recommendation method and a system for recommending a graph neural network in a cross-border mode. Background Nowadays, the graphic neural network is widely used for commodity recommendation systems. In the cross-border recommendation scene, due to the fact that local map data has the problem of sparse data, enough training samples are lacked to train a cross-border recommendation model. In order to provide enough training data, a cross-border recommendation model is generally issued to clients located around the world through a federal learning framework, the clients train by using local map data, and model updates uploaded by the clients are aggregated at a server to form a global model. However, the graph structure constructed by each client has systematic differences in structure, semantics, and dynamics, though overlapping in feature space, due to factors such as regional culture, policy restrictions, consumption habits, and the like. The systematic difference causes significant difference in distribution of the local model of each client in the parameter space, so that the parameters of the global model are difficult to effectively converge in the same region during aggregation, and the prediction accuracy of the global model is not obviously improved. Disclosure of Invention The embodiment of the application mainly aims to provide a model training method, a recommendation method and a system for a cross-border recommendation graph neural network, so that a cross-border recommendation model of a client can aggregate information from local graph data through a new aggregation paradigm, and each client forms a cross-border recommendation model with overlapped distribution in a parameter space, so that global model parameters aggregated at a server can be converged. In order to achieve the above objective, an aspect of the embodiments of the present application provides a model training method of a cross-border recommendation graph neural network, applied to a client of federal learning, the method including: determining an object interaction sequence based on the historical interaction record of the target object, and determining an item interaction sequence based on the historical interaction record of the target item; Constructing a dynamic bipartite graph through the object interaction sequence and the project interaction sequence, wherein the dynamic bipartite graph comprises bipartite graphs corresponding to N time steps, and N is an integer greater than or equal to 1; constructing or updating a cross-border recommendation model, wherein in the process of transmitting information on the dynamic bipartite graph, a first embedded vector of a second-order item neighbor of the target object is aggregated, the second-order item neighbor is connected with the target object through an implicit association object, the first embedded vector represents a favorite item of the target object, a second embedded vector of the second-order item neighbor of the target object is aggregated, the second-order object neighbor is connected with the target item through an implicit association item, and the second embedded vector is directed to represent an attractive object of the target item; training the cross-border recommendation model based on the training triples and the dynamic bipartite graph, and sending model update data to a server, so that the server determines a global model according to the cross-border recommendation models of a plurality of clients. In some embodiments, constructing a dynamic bipartite graph through the object interaction sequence and the item interaction sequence specifically includes: converting each interaction behavior in the object interaction sequence into an object dense vector, and embedding the object dense vector according to time attenuation to obtain an embedded representation of the target object; Converting each interaction behavior in the item interaction sequence into an item dense vector, and embedding the item dense vector according to time attenuation to obtain an embedded representation of the target item; And constructing the dynamic bipartite graph according to the embedded representation of the target object and the embedded representation of the target item. In some embodiments, the cross-border recommendation model is provided with a GNN module and a perception module; the training of the cross-border recommendation model based on the training triples and the dynamic bipartite graph specifically comprises the following steps: Encoding the training triples through interaction scoring tags, and determining an adjacency matrix through the dynamic bipartite graph; The method comprises the steps of inputting a training triplet afte