Research On Classification Of Online Product Reviews’ Helpfulness Based On Multi-feature Fusion

With the rapid development of e-commerce,online shopping users are increasing with each passing day.As an important part of e-commerce platform,online reviews provide users with additional commodity information and assist users to make decisions.However,a large number of users make the number of online reviews grow explosively,leading to the relatively helpful reviews are difficult to be discovered.Therefore,it is of great practical value and significance to study how to accurately and efficiently judge the helpfulness of reviews.At present,domestic and foreign researches on the helpfulness of reviews mainly focus on the influencing factors of helpfulness,that is,the construction and selection of the reviews’ helpful features,which mainly includes two categories.One is the external features constructed by statistical analysis of data sets as the attached attributes of reviews.The other is the textual features constructed by using the review text as the subject of the review.Through research and analyse,external features and textual features are important influencing factors for judging the helpfulness of reviews,and they should be considered together.Based on the relevant research on the factors influencing the helpfulness of reviews,from the viewpoints of review,reviewers,commodities,constructs the review length,review score,score deviate,readability,review relevance,reviewer’s credibility,reviewer’s experience and time deviate,altogether eight external features and the textual features.A multi-feature fusion model was constructed by combining the external features with the textual features after dimension reduction through ensemble learning algorithm,which fully considered the effect of external features and the textual features,at the same time,avoided the imbalance of the number of features.In order to verify the effect of the model,two groups of control experiments were constructed in this dissertation.The first group included the Bert combined multi-layer neural network classification model experiment,the simple feature fusion classification model experiment,the multi-feature fusion classification model experiment and the TextCNN classification model experiment.The second group included the PCA classification model experiment and the LDA classification model experiment.According to the experimental results of the two groups,the multi-feature fusion classification model has a better effect,indicating that the hierarchical fusion of external features and review textual features can effectively improve the effectiveness of recognition of reviews’ helpfulness,and the text feature dimensionality reduction combined with Bert multi-layer neural network can better extract text information and further improve the accuracy of the model.