Research On Helpfulness Prediction And Impact Factors Of Online Product Reviews

With the rapid development of the Internet, the business of network has been popular and network consumption has become significant in the daily life of people. However, due to the uncertainty of the products in the network transactions, it is urgent to get information to help individuals for purchase decisions and to help enterprises for feedback collection. Therefore, online reviews have received more and more attention. However, it is difficult to obtain useful information for us due to a large number of invalid and even malicious reviews, thus, how to get really useful reviews in the huge comments attracts more attention of researchers.This dissertation focuses on how to predict the usefulness of online reviews and main contributions are as follows.(1) We give a survey of predictive review’s usefulness:More specifically, we first introduce the background and motivation, and then summarize some methods of predictive review’s usefulness. Finally, we present our approach in the analysis of these methods.(2) We build a RRS-L model and give the analysis of impact factors:Existing efforts mainly focus on the analysis of reviews’ text property. Motivated by this, we propose a multiple linear regression algorithm based method to predict the helpfulness of online reviews. The method first considers three impact factors including reviews’ text properties, reviewers’ properties and stores’ properties, then it creates a RRS-L model to predict the helpfulness of online reviews. Experimental results conducted on real data sets show that the model has a good classification effect, can effectively filter useless reviews, but also can eliminate the useless independent variables.(3) Redundancy analysis of independent variables in RRS-L model:In the analysis of the impact factors, we get the RRS-LL model. To ensure the stability of the model, we analyze the linear correlations among independent variables in the RRS-L model, and remove the independent variables with strong linear correlations. Finally, we build models as variables arrive one by one and get an optimal model called RRS-LL Experimental results show that the RRS-LL model can perform better compared to the classical two classification models, while using fewer independent variables and reducing the cost.