Research On Interpretable Book Recommendation Based On Knowledge Graph

With the advent of the big data era,recommendation systems have become important technologies widely applied.However,traditional recommendation systems have problems in terms of interpretability and data sparsity.This thesis aims to address these issues by proposing a knowledge graph-based interpretable book recommendation method.By analyzing multiple auxiliary resources such as reader book reading records,book information,reader comments,reader ratings,and sentiment information,the reader’s interests and preferences are obtained.The method combines reinforcement learning inference to generate recommendation paths,improving recommendation accuracy while making the results interpretable,which is of great practical significance for improving reader satisfaction and book sales on online book platforms.This thesis mainly explores the massive data of reader-book interactions on the Douban book platform and conducts in-depth research on online comments to achieve the goal of interpretable book recommendation using reinforcement learning and knowledge graph technologies.To achieve this goal,a crawler program is used to collect a large amount of data,and methods such as knowledge graph,sentiment analysis,and reinforcement learning are employed for research.The research includes three aspects:(1)Constructing a book knowledge graph by combining the Neo4 j graph database.A network crawler program is used to collect real and effective reader-book interaction data from the Douban book platform,including reader comment texts.After pre-processing steps such as deduplication,default normalization,and traditional Chinese to simplified Chinese conversion,the term frequency-inverse document frequency(TF-IDF)algorithm is used to extract feature words of comments to identify the reader’s interests and book features.Finally,the book knowledge graph is constructed using knowledge graph construction methods and stored in the Neo4 j database,laying the foundation for the construction of the book sentiment knowledge graph.(2)To accurately classify the emotional information of book comments,an improved Enhanced Representation through Knowledge Integration(ERNIE)sentiment classification model is proposed.The improved ERNIE pre-training model is used to perform sentiment analysis on the experiment data comment set of Douban books.By adding a fully connected layer,the performance of the ERNIE pre-training model on the dataset is improved.Experimental results show that using the improved pre-training model can improve the accuracy and precision of sentiment classification tasks.(3)Knowledge graph-based interpretable book recommendation research.Different from existing methods,this method aims to generate interpretable causal reasoning processes through explicit reasoning,rather than just focusing on using knowledge graphs to obtain more accurate recommendations.By providing interpretable paths in the sentiment knowledge graph,recommendation and interpretability are combined.Firstly,the improved ERNIE pre-training model is used to perform sentiment analysis on reader comment texts.Then,the results of the sentiment classification model are weighted with the reader’s ratings to form three types of emotional relationships: interested,generally like,and dislike,which are added to the book knowledge graph to construct the book sentiment knowledge graph.Secondly,a reinforcement learning method based on Markov decision is used to infer the recommendation path in an interpretable way by developing policy functions and training intelligent agents.Finally,compared with multiple advanced benchmark algorithms,the experimental results show that the proposed recommendation algorithm outperforms the benchmark algorithms in terms of normalized discounted cumulative gain,accuracy,hit rate,and recall rate.Moreover,it can provide readers with interpretable recommendation processes,making the recommendation results more convincing and providing an effective solution for book recommendation.