Analysis of Customer Journeys Using Prototype Detection and Counterfactual Explanations

This paper introduces a novel three-step approach for analyzing **customer journeys**, which are sequences of steps consumers take during interaction with a product or brand. Traditional methods struggle with the sequential and complex nature of customer journey data. The proposed method first calculates the distance between these sequences using a weighted Levenshtein distance tailored for staged data, which allows them to **find and visualize typical journey patterns** using clustering (k-medoids) and visualization (MDS). Second, based on these calculated distances, they predict the outcome, specifically **whether a customer will make a purchase**, using a method like k-nearest neighbors (k-NN). Finally, if a journey sequence suggests no purchase will occur, they propose a method to extract **"counterfactual" sequences** from the existing data that are similar but resulted in a purchase, thereby suggesting actionable changes to improve the non-purchasing journey. The approach was evaluated using survey data about cosmetic purchases, demonstrating its ability to extract interpretable patterns, predict purchases with reasonable accuracy, and identify which changes in a journey could influence buying decisions, ultimately aiming to support improved marketing activities. https://arxiv.org/pdf/2505.11086