Date of Award
Doctor of Philosophy
Sanjoy Ghose, Amit Bhatnagar, Xiang Fang, Zuhui Xiao
deep learning, fake review detection, hotel review rating, machine learning, managerial response
This dissertation investigates how online reviews and managerial responses jointly affect consumer perceptions. I first examine and compare the outcomes of multiple fake review classifiers using various algorithms, including traditional machine learning methods and recently developed deep learning methods (essay I). Then, based on the findings of the first essay, I examine the interrelationship between fake review detection, managerial response, and hotel ratings and ratings’ growths (essay II).The first essay is a comparative study on the methodology of identifying fake reviews. Although online reviews have attracted much attention from academia and industry for over fifteen years, how to identify fake reviews is still under study. In terms of methods, traditional machine learning classification methods were in dominant use. Recently, with the rise of deep learning methods in text analysis since the 2010s, researchers began applying new deep learning classification methods. In terms of features, the way to extract information from review content has been developing as the Natural Language Processing (NLP) area has made much progress since 2013. After that, researchers tried to apply both deep learning algorithms and extract dense text features to build alternative systems for identifying fake reviews. Among various algorithms and features, how to choose and set up a good fake review detector, demands researchers to explore further to arrive at a widely accepted answer. This study is the first that applies both traditional machine learning and deep learning methods and compares across multiple datasets that vary in size, origin, and class distribution. This paper reports three findings. First, with new deep learning algorithms, classifiers perform better than classifiers using traditional machine learning methods in most cases, with only a few exceptions. Second, with dense word embeddings, classifiers perform better than classifiers using one-hot text features. Third, incorporating other numerical features boosts classification performance. The purpose of the second study is twofold. First, to explore factors contributing to the likelihood of a review to receive a managerial response (MR), testing the impact of the fake review detection results, review congruency, review deviation, and the moderating role of hotel class. Second, to examine the association among online reviews, managerial responses, and hotel rating (growth rate), including both text similarity and fake review detection results as independent variables. This study is one of the first that introduces fake review detection and text similarity into research about MR, adding to the literature of MR in the context of Tripadvisor.com. Our findings indicate the following practical implications. (1) A truthful, detailed, and congruent review is more likely to receive an MR; (2) The percentage of truthful reviews has a strong and positive association with hotel rating and its growth. In an extreme situation, the hotel rating will go up by 0.21, and the rating growth rate will increase by 8.5% due to 100% truthful reviews; (3) Hotels should carefully choose which review(er)to respond to and make responses concise and matching while actively responding to reviews.
Chen, Long, "Essays on Fake Review Detection, Managerial Response, and Consumer Perceptions" (2021). Theses and Dissertations. 2653.