Amazon Phone Reviews Analysis: Using an Econometric and Deeplearning Approach
Sebastián Chamarravi*, Daniel Duque and Bilal Faraj
ABSTRACT
In this study, 334,328 comments on unlocked mobile phones from Amazon are analyzed, exploring descriptive statistics and constructing predictive models for rating and price using machine learning techniques, an analysis that provides information still lacking sufficient robustness in the literature. To do this, reviews were processed to generate an approximate sentiment score using the TextBlob library, related to the user's rating. In turn, simple and multiple linear regressions, and regressions including interactions, were performed to predict price, as well as multiple linear regression models to predict rating using sentiment score, price, and Review Votes (support for other users' reviews).
Additionally, Multilayer Perceptron (MLP) neural network models and Transformer based models from the Hugging Face library were trained for both prediction tasks. The results show a moderate correlation between sentiment score and Rating (r≈0.554) and a significantly higher fit for rating prediction with more complex models (R²≈0.727) than with linear regression (R²≈0.308). In contrast, all approaches were poor at attempting to predict price (R²≤0.086), indicating the limited linear relationship between the available attributes and the product's price.
