Author

Scott Bolter

Publication Date

Spring 2013

Degree Type

Master's Project

Department

Computer Science

Abstract

In this paper we focus on automatically classifying product reviews as either helpful or unhelpful using machine learning techniques, namely, SVM classifiers. Using LIBSVM and a set of Amazon product reviews from 25 product categories, we train models for each category to determine if a review will be helpful or unhelpful. Previous work has focused on training one classifier for all reviews in the data set, but we hypothesize that a distinct model for each of the 25 product types available in the review dataset will improve the accuracy of classification. ! Furthermore, we develop a framework to inform authors on the fly if their review is predicted to be of great use (helpful) to other readers, with the assumption that authors are more likely to rethink their review post and amend it to be of maximum utility to other readers when given some feedback on whether or not it will be found helpful or unhelpful. ! Using past research as a baseline, we find that specialized SVM classifiers outperform higher level models of review helpfulness prediction.

Share

COinS