Predicting Customer Satisfaction by Amazon Reviews: The Bluetooth Earphones Example

The present study focuses on how to predict customer satisfaction from textual data related to customer feedback, by means of a case study regarding a collection of 92,000 Amazon reviews referring to Bluetooth earphones for listening to music. We exploit Semi-supervised Latent Dirichlet Allocation to extract from the reviews the most relevant topics by using two different ad hoc dictionaries, thus deriving useful information about the competitive positioning of the considered brands. In the following, we explicitly model star ratings by means of the extracted topics, using a logistic regression and a classification tree approach. We employ both the whole sample and a selection of 1- and 5-star rating reviews (which boosts accuracy by over 77%). In this way, we manage to rank extracted topics by their predictive value of customer feedback, which is potentially very useful to improve customer satisfaction in a text-driven way.