Personalized Review Ranking for Improving Shopper's Decision Making: A Term Frequency based Approach

TL;DR

This paper introduces a personalized review ranking method that leverages user profiles and term frequency analysis to improve shopper decision-making and product recommendations, addressing the challenge of relevance variability without explicit labels.

Contribution

It proposes a novel personalized review ranking approach using user profiles and term frequency, enhancing review relevance and product recommendations without labeled data.

Findings

Significant improvement in top review quality and user satisfaction.

Enhanced product recommendation accuracy based on review texts.

Effective personalization without explicit relevance labels.

Abstract

User-generated reviews serve as crucial references in shopper's decision-making process. Moreover, they improve product sales and validate the reputation of the website as a whole. Thus, it becomes important to design reviews ranking methods that help shoppers make informed decisions quickly. However, reviews ranking has its unique challenges. First, there is no relevance labels for reviews. A relevant review for shopper A might not be relevant to shopper B. Second, since shoppers cannot click on reviews, we have no ways of getting relevance feedback. Eventually, reviews ranking suffers from the lack of ground truth due to the variability in the standard of relevance for different users. In this paper, we aim to address the challenges of helping users to find information they might be interested in from the sea of customer reviews. Using the Amazon Customer Reviews Dataset collected and organized by UCSD, we first constructed user profiles based on user's personal web trails, recent shopping history and previous reviews, incorporated user profiles into our ranking algorithm, and assigned higher ranks to reviews that address individual shopper's concerns to the largest extent. Also, we leveraged user profiles to recommend products based on reviews texts. We evaluated our model based on both empirical evaluations and numerical evaluations of review scores. The results from both evaluation methods reveal a significant increase in the quality of top reviews as well as user satisfaction for over 1000 products. Our reviews based recommendation system also suggests that there's a large chance of user viewing and liking the product we recommend. Our work shows the basic steps of developing a ranking method that learns from a particular end-user's preferences.