Automated Techniques for estimating Customer Value and Causal Models of Customer Satisfaction

Abstract

In the competitive business environment of today, knowing which customers to target,and how, is essential for maximizing the return of the invested resources.In order to accomplish this we need a way of estimating who will be the most valuablecustomers at a firm, as well as a method to increase the probability that the customersreturn. We analyze the extended Pareto/NBD model’s performance at predicting futurevaluable customers, and compare it to a simple heuristic as well as the followingmachine learning techniques: the M5' model tree and the multilayer perceptron (neuralnetwork). Furthermore, we apply causal network algorithms to customer satisfactionsurvey data to identify causal relations and how they drive customer satisfaction. Wecompare this to the results of a correlation based approach. We test the Fast CausalInference Algorithm (FCI), Greedy Equivalent Search (GES) and the IC / PC algorithm, ofwhich the last two are Bayesian Network algorithms.Our results show that the extended Pareto/NBD model performs at the same level asthe M5’ model tree and the multilayer perceptron. Furthermore, the extendedPareto/NBD model estimates the aggregated spending of the customer bases from ourtest data with a worst case error of 27 percent. For the causal networks algorithm wefind that the networks learned by the different algorithms give additional informationabout the relationships in the data when compared to a pure correlation approach. Forthe PC and GES algorithms we observe that the questions from different aspects of thehotel experience cluster in an appropriate manner. However, the validity of some of therelationships detected by the algorithms is disputable.Finally we describe how a customer value estimation model, such as the extendedPareto/NBD model, can be supplemented with a causal network algorithm to form aframework for estimating potential revenues as a result of improving various aspects ofa business. While the framework is still at an idea stage, it shows the potential of areliable causal network algorithm, motivating further research in the area.