Rule Learning Model Based On E-commerce Knowledge Graph

Rule learning is an important field with broad application prospects in data mining For the task that generates rules from e-commerce knowledge graphs with specific data structures,widely-utilized association rule mining algorithms,although simple and understandable,cannot be used to mine association rules for data with specific data structures.At present,most of the rule learning algorithms based on knowledge graphs are aimed at dense knowledge graphs,and their characteristics of graph traversal-based search also make such algorithms suffer from efficiency problems on large-scale knowledge graphs.The representation learning methods used for knowledge graph reasoning mostly focus on the inference results and cannot generate high-quality rules that can be easily understoodTherefore,focusing on the task of generating specific rules in the field of e-commerce,this article,from the perspective of knowledge graph,constructs the existing property item P(Property)and property value V(Value)and their corresponding scenarios into a relation-specific knowledge graph,at the same time,structures the mounting problem as a link prediction problem.With the utilization of attention mechanism,a rule learning model based on knowledge graph representation learning is proposed in this article,the weight of the product attributes is obtained by the representation of the scene and mounting relationship.This model can generate rules from existing data to mount new products,improve the efficiency of product mounting,and reduce labor costsAlso,from the perspective of reinforcement learning,this article constructs the product mounting problem as a classification problem and the rule generation problem as a sequence decision problem,as a result,a rule learning model based on reinforcement learning is proposed which trains an agent to generate the body part of rules.In the terms of length of rule's body part,a rule length term is added to the reward function to generate a shorter rule.And for the reason that the length of the product data sequence is long,the agent needs to explore a larger sequence decision space,in order to allow the model to explore a larger space,this article also applies methods including Rollouts,Reward Baseline,and adding information entropy to help model training,and achieves good results.