Detecting Pain Points from User-Generated Social Media Posts Using Machine Learning

Abstract

Artificial intelligence, particularly machine learning, carries high potential to automatically detect customers’ pain points, which is a particular concern the customer expresses that the company can address. However, unstructured data scattered across social media make detection a nontrivial task. Thus, to help firms gain deeper insights into customers’ pain points, the authors experiment with and evaluate the performance of various machine learning models to automatically detect pain points and pain point types for enhanced customer insights. The data consist of 4.2 million user-generated tweets targeting 20 global brands from five separate industries. Among the models they train, neural networks show the best performance at overall pain point detection, with an accuracy of 85% (F1 score = .80). The best model for detecting five specific pain points was RoBERTa 100 samples using SYNONYM augmentation. This study adds another foundational building block of machine learning research in marketing academia through the application and comparative evaluation of machine learning models for natural language–based content identification and classification. In addition, the authors suggest that firms use pain point profiling, a technique for applying subclasses to the identified pain point messages to gain a deeper understanding of their customers’ concerns.