A Trivariate Customer Satisfaction Prediction Framework Using XLNetBased Text Mining Models

Abstract

Customer satisfaction prediction has become increasingly critical as organizations receive thousands of support tickets daily, with over 80% of customers basing loyalty on timely and accurate service responses; therefore, real-time text analytics plays a vital role in transforming raw customer messages into actionable insights. This research focuses on a trivariate classification framework that simultaneously predicts ticket priority, customer satisfaction categories, and likelihood of issue resolution; however, traditional single-task approaches fail to capture the interconnected nature of these three dimensions, creating the need for an integrated model. The problem arises because many organizations still rely on manual methods for customer satisfaction prediction, including basic text mining and rule-based similarity checks, which are slow, inconsistent, and unable to handle largescale unstructured data effectively. The objective of this study is to develop an XLNet-driven model combined with machine-learning methods such as Quadratic Discriminant Analysis (QDA), Linear Discriminant Analysis (LDA), Stochastic Gradient Descent (SGD), and Nearest Centroid (NC), but these traditional classifiers produce lower accuracy, whereas the Histogram-Based Gradient Boosting (HGB) approach significantly enhances predictive performance.