Metaheuristic-Based Machine Learning Approach for Customer Segmentation

Abstract

In the globalized knowledge economy, the challenge of translating best available evidence from customer profiling and experience into policy and practice is universal. Customers are diverse in nature and require personalized services from financial institutions, whereas financial institutions need to predict their wants and needs to understand them on a deeper level. Customer segmentation is a very crucial process for a financial institution to profile new customers into specific segments and find patterns from existing customers. Usually, rule-based techniques focusing on specific customer characteristics, according to expert knowledge, are applied to segment them. However, these techniques highlight the fact that traditional classifications in the big data era are becoming increasingly irrelevant and agree to the claim of financial institutions not knowing their customers well enough. The main objective of this work is to propose an evolutionary clustering approach as a rule extractor mechanism that facilitates decision makers to recognize the most significant customer characteristics and profile them into segments. Particularly, a population-based metaheuristic algorithm (Genetic Algorithm) is used in a hybrid synthesis with unsupervised machine learning algorithms (K-means Algorithms) to solve data clustering problems. Based on the clustering result, labels are added for every data point in the dataset. This dataset is used to train supervised ML algorithms such as deep learning and random forests to predict in which cluster a new customer can be mapped. A cluster analysis conducted on behalf of the EXUS financial solutions company that provides financial institutions with financial software that can deliver debt collection services effectively, meeting both academic requirements and practical needs. Two real-world datasets collected from financial institutions in Greece, explored and analyzed for segmentation purposes. To demonstrate the effectiveness of the proposed method, well-known benchmark datasets from UCI machine learning repository were also used.