On Inferring a Meaningful Similarity Metric for Customer Behaviour

Abstract

In omnichannel customer service environments, where no real process is enforced, a wide variety of customer journey variants exists. This variety makes it complex to find process improvement opportunities. Modeling the journeys as traces is an essential step before discovering an explainable model of various behaviours. Trace clustering helps improvement efforts by separating the journeys into homogeneous subsets in terms of behaviour and purpose. For this, a one-size-fits-all distance metric has been used so far in the literature. This paper shows that a domain-informed similarity metric will improve customer journey clustering compared to a generic one. We propose SIMPRIM framework, which uses clustering quality metrics to develop a similarity metric that maximizes the separability of the journeys in a low dimensional space while agreeing with existing process knowledge. Experimental evaluation on real life use cases of a large telecom company and a benchmark dataset show that, compared to a generic metric, respectively a 46% and 39% improvement can be obtained in terms of the internal clustering quality while keeping the external clustering quality equal. We also show that the inferred metric can be useful for prediction applications.