Analysis Methods Of Inferring Complementary And Substitutable Products Based On Knowledge Graph

With the rapid development of the Internet,the number of products on e-commerce platforms continues to grow,and the relationships between products have become more complex and indistinguishable.Understanding the complementary and substitution relationships between products can not only improve the accuracy of e-commerce recommendations,but also help retailers to make better operation and management decisions,such as bundle sales,promotion and pricing,warehouse goods layout and so on.However,in the scenario of large-scale data in e-commerce,the product relationship analysis faces the following two challenges:(1)The heterogeneity of diverse data increases the difficulty of data fusion and feature extraction;(2)The interpretability appeal of product relationships increase the complexity of problem handling.Knowledge graphs,thanks to their specific network structure,can fuse products and product-related data with each other to create more comprehensive semantic relationships between entities,while enabling interpretable reasoning results.Therefore,this thesis investigates the analysis method of inferring complementary and substitutable products based on knowledge graphs,and the main work is as follows:(1)A knowledge graph has been constructed and a knowledge representation learning model that incorporates entity descriptions has been proposed.Firstly,the structured data in the diverse data is extracted to construct the knowledge graph.Then,the triad structure information and entity description information in the knowledge graph are embedded using Trans E model and SBERT model respectively,and the two learned knowledge representations are integrated together,and finally the fused knowledge representations are trained by Trans E model to obtain the final knowledge representation.(2)A Markov Decision Process-based inference model for product relations has been constructed and a reward function has been improved based on cosine similarity.First,a product relationship inference model based on Markov Decision Process is built,and the cosine similarity of the inference path and query relationship is proposed to be added to evaluate the accuracy of the inference path,which feedbacks to the intelligence as a reward function to train the optimal policy network.Then,the actions guided by the policy network are selected by beam search,and the relational inference results are obtained by ranking the scores of all commodity nodes in the obtained path network,while the interpretation paths are given as well.(3)Numerical experiments and analysis are deployed.Numerical experiments are conducted on the typical e-commerce platform dataset.The effectiveness of the present model in the product relationship inference task is demonstrated by validity experiments compared with other baseline algorithms.The effectiveness of the improved knowledge representation learning model and reward function in this thesis is demonstrated through ablation experiments.The effects of these parameters on the model performance are illustrated by sensitivity experiments.Through interpretability analysis,it is demonstrated that the model is able to provide an explanatory path for the inference results.In summary,in order to address the data fusion and interpretability problems faced in the task of product relationship inference for data,the methods proposed in this thesis can solve the problem of difficult feature extraction from diverse data,improve the relational inference capability of the model,and make the inference results interpretable.In addition,the product relationship and the explanation path obtained by the relational inference model can help the implementation of the recommendation system,can provide a basis for decision making on ecommerce bundle sales,promotion and pricing,and warehouse goods layout,which is of great practical value.