Iterative Threshold-based Naïve Bayes Classifier: an efficient Tb-NB improvement

Romano, Maurizio

First

;Zammarchi, Gianpaolo

Second

;Contu, Giulia

Last

2022-01-01

---

Abstract

While analyzing online reviews on Booking.com, we proposed an ad-hoc classification model (Threshold-based Naïve Bayes Classifier, Tb-NB) to evaluate Customer Satisfaction, starting from the reviews' content, and predicting them as positive/negative. The log-likelihood ratios attributed to each word included in a review are then used to estimate a numeric sentiment score. In this paper we propose an improved version of Tb-NB called "iterative" Tb-NB. It results in a second step of Tb-NB: starting from the output of Tb-NB and reclassifying reviews with a probabilistic approach, it refines iteratively the threshold value used to classify a given subset of reviews.