Predicting customer satisfaction with product reviews: A comparitive study of some machine learning approaches.

Abstract

In past two decades e-commerce platform developed exponentially, and with this advent, there came several challenges due to a vast amount of information. Customers not only buy products online but also get valuable information about a product they intend to buy through an online platform. Customers share their experiences by providing feedback which creates a pool of textual information and this process continuously generates data every day. The information provided by customers contains both subjective and objective text that contains a rich information regarding behaviour, liking and disliking towards a product and sentiments of customers. Moreover, this information can be helpful for the customers who are yet to buy or who are yet in decision making process. This thesis studies comparison of four supervised machine learning approaches to predict customer satisfaction. These approaches are: Naïve Bayes, Support Vector Machines (SVM), Logistic Regression (LR), and Decision Tree (DT). The models use term frequency inverse document frequency (TF-IDF) vectorization for training and testing sets of data. The models are applied after basic pre-processing of text data that includes the lower casing, lemmatization, the stop words removal, smileys removal, and digits removal. We compare the performance of models using accuracy, precision, recall, and F1-scores. Support Vector Machines (SVM) outperforms the rest of the models with the accuracy rate 83% while Naïve Bayes, Logistic Regression (LR) and Decision Tree (DT) have accuracy rate 82%, 78%, and 76%, respectively. Moreover, we evaluate the performance of classifiers using confusion matrix.