Algorithm And Application Of A Knowledge Graph Enabled Recommendation System For Spare Parts Vertical E-commerce

With the development of China’s e-commerce,many spare parts manufacturer have gradually established their own vertical e-commerce platforms and conducted online ecommerce.However,the number of types of sapre parts,also known as SKU（stock keeping unit）,on the platform is large,and it would take a lot of time for users to find suitable spare parts.If users’ interested objects can be accurately identified and recommended by algorithm,the user experience will be greatly improved and greater platform benefits will be achieved.However,there are two problems in the application of traditional recommendation algorithms in vertical e-commerce.First,different from general e-commerce platforms,the data on vertical ones is highly sparse,leading to insufficient data for modeling and even low prediction accuracy.Second,the association relationships among items are not fully explored,making poor interpretability of recommendation results.These problems widely exist in vertical ecommerce platforms and thus worth investigation.Characteristics of the spare parts vertical e-commerce platform provide an opportunity to solve these problems.First,the platform has numerous user behavior data where the general platform may lack.Many vertical platforms have established groups such as designers,followers,and chatting.Users actively publish works and initiate discussions in these groups.These rich behavioral data are not available in general platforms,and also provide a large amount of personalized data to further improve the accuracy of recommendation algorithm.Second,the platform has a large amount of relationship information on items.The spare parts of the platform are not only independent commodities,but also can be combined with other spare parts to form assemblies.By constructing and structuring the combination relationship among various spare parts,and mining knowledge and rules,we are able to provide an effective solution to enhance the interpretability and robustness for the spare parts vertical ecommerce recommendation system.Therefore,this thesis proposes a vertical e-commerce recommendation algorithm based on the knowledge graph of spare parts and user behaviors,and applies and validates the algorithm in a cooperated company,including the following work.First,we conducted data extraction,preprocessing,feature construction,and storage from the studied vertical platform,and then constructed ontology representation,knowledge extraction,and graph storage with the cleaned data.Based on this work,we completed the definition and construction of the knowledge graph of spare parts and user behaviors for the studied vertical platform.Second,we built a recommendation model that can capture user preferences from the knowledge graph.To incorporate node attribute information into the knowledge graph,a streaming node embedding structure is proposed,which can fully mine and utilize node attribute information.Aiming to reduce time complexity in node traversing,an important node sampler is proposed,which can substantially reduce noise and improve efficiency in node attribute processing.To examine the performance of the proposed method,this thesis compares and analyzes the new model with the conventional ones.Results show that the proposed model outperforms other four models.The accuracy and AUC are improved by 10.2%and 16.0% compared with the previous best model.Third,to explore the dynamic patterns of our knowledge graph,and to investigate the influence of graph variations on the performance of the recommendation system,this thesis novelly conducts time series analysis and mutation analysis on the topological features of the knowledge graph,and determines the optimal form and shape of the knowledge graph structure.In addition,in order to reduce the cost of updating the knowledge graph in real time when the model is actually applied,this thesis uses the graph topology index to predict the recommendation accuracy,and determines the optimal graph update interval should be 10 days,which can effectively reduce the cost of platform computing resources.Finally,a personalized recommendation system for spare parts vertical e-commerce is designed and implemented in the field of vertical e-commerce of spare parts,with the personalized recommendation algorithm proposed in this thesis.Based on user historical interaction data and relationship information on building block spare parts,we accurately recommend building blocks for users in the cooperated vertical e-commerce platform.In this thesis,online A/B test is carried out on the platform as well.Results show that the recommendation success rate has increased by 25.28%,suggesting the effectiveness of the proposed algorithm in real application scenario.