Big Data Analytics for Reputational Reliability Assessment Using Customer Review Data
Résumé
Traditionally, reliability assessment is done based on lifetime testing data. Such assessment methods suffered from a lot of limitations. For example, it is in general difficult to collect enough life testing data to support an accurate reliability assessment. Further, the experimental conditions can hardly reproduce the way a consumer will use a product in practice. In the meantime, with the expansion of the Internet, a lot of customers give their feedbacks on the products by posting reviews on websites. This constitutes a huge, easily accessible, and more realistic database that can be used to assess reliability. In this work, we scraped reviews from a famous e-commerce website. Machine learning models are developed to extract failure-related information from these reviews. Two kinds of information are examined in this study: (1) whether a review reports a failure and, in such a case, (2) the severity of the failure. We used natural language processing tools to process text and we developed different classification models for information extraction. The developed methods were tested on customer review data from 11 different tablets of several brands. The results we obtained were around 85% accuracy when training and testing our models with our dataset. Hence, the machine learning-based approach we developed is demonstrated to be a promising first step to assess reliability thanks to web-based data. However, with a corpus containing only a few thousand reviews and more than 100,000 words, using text to train classification models remains a complicated task. Especially, the models developed in this paper strongly overfit despite the use of several methods designed to prevent overfitting.