Hinal Jajal

Product reviews form a goldmine of text data that is helpful to both consumers and businesses in making buying decisions and understanding consumer needs respectively. In this study, we present a method to convert review text data into a product-reviewer network that can be used to predict consumer sentiment towards a product they have not previously reviewed. Our method utilizes two key techniques in this process: the first is natural language processing to extract the sentiment from thousands of reviews and the second is balance theory to analyze the signed network. We address the challenges in drawing insights from a signed, bipartite network and make use of existing literature on modified balance theory to overcome them. With a carefully sampled dataset of Amazon fashion products’ reviews, we demonstrate the effectiveness of our method in predicting the signs of consumer sentiment. We explore three sign prediction methods, namely the Signed Caterpillar (SCsc), Random Walk Based Models (SBRW), and Matrix Factorization (MFwBT), and successfully implement two of the three with remarkable accuracy. We recognize the several limitations in our processes, including discrete sentiment values and a limited dataset. Future work would expand these promising results into a more comprehensive study addressing the mentioned limitations.

Detailed study completed for the Math 168 class at UCLA can be found here.