Failure Mode Extraction via Vectorization and Clustering: A Case Study on Amazon Reviews

Abstract

Due to the fact that consumers have mostly preferred e-commerce sites as a purchasing channel in recent years, it is necessary to focus on the online marketing processes for the manufacturers and get more benefit from it. Thus, product reviews on online sites provide great feedbacks for companies to improve their products or services and for potential customers on purchase decisions. However, it is not possible to evaluate every single feedback for a specific product by a company, due to the immense amount of post-sale reviews for a certain product. In that study, negative categorical customer reviews are crawled for a specific product from amazon.com, then reviews are converted into vectors using Doc2Vec and finally, reviews are clustered by using k-means, and DBSCAN algorithms. As a result of clustering and cluster labeling, the most frequent topics (failure modes) are extracted along with pointless reviews. The result of the clustering presented useful managerial insights for product improvement processes and a beneficial perspective for potential customers on purchasing decisions concerning the topic frequency and mean star values. Besides, the proposed pipeline can be used as a topic modeling approach in terms of its promising results. © 2024 Elsevier B.V., All rights reserved.