Modernising customer service in retail: A Worten case study on automated complaint classification

Abstract

The emergence of retailers able to deliver products on the same day and at very competitive prices, such as Amazon, has caused customers to raise their expectations. When the quality of service falls short of the expected, customers resort to complaints to show their dissatisfaction, and it is in the retailers' interest to resolve the problem as quickly as possible to avoid losing customers. Since the process of analysing complaints is very time-consuming, this study aims to propose a method for classifying the complaints addressed to Worten, automatically. Thus, sixteen experiments were performed with eight different Machine Learning (ML) algorithms, following the Cross Industry Standard Process for Data Mining (CRISP-DM) methodology. The experiments included reducing the number of classes, Transfer Learning models, and different types of class balancing, among others. The Support Vector Machine (SVM) model obtained the best classification, with an Accuracy of 71.41%, in the experiment in which the three most diffuse of the six original classes (Time, Technical Problem, Client, Money, Service and Other) were eliminated.

O aparecimento de retalhistas com capacidade de entregar produtos no próprio dia e a preços muito competitivos, como a Amazon, levou a um aumento de expectativas por parte dos clientes. Quando a qualidade do serviço praticado fica aquém do esperado, os clientes recorrem a reclamações para demonstrar o seu descontentamento, e é do interesse dos retalhistas resolver o problema o mais rápido possível para evitar perder clientes. Uma vez que o processo de análise de reclamações consome bastante tempo, este estudo visa propor um método de classificar as reclamações endereçadas à Worten, de forma automática. Assim, foram realizadas dezasseis experiências com oito algoritmos de Machine Learning (ML) diferentes, seguindo a metodologia Cross Industry Standard Process for Data Mining (CRISP-DM). As experiências efetuadas compreenderam a redução do número de classes, modelos de Transfer Learning, diferentes tipos de balanceamento de classes, entre outros. O modelo Support Vector Machine (SVM) obteve a melhor classificação, com uma Acurácia de 71,41%, na experiência em que foram eliminadas as três classes mais difusas das seis classes originais (Time, Technical Problem, Client, Money, Service e Other).