Customer satisfaction is a key performance indicator, usually measured and analyzed using retrospective surveys. Retrospective analysis, however, offers limited value since the organization cannot use it to address problems as they arise. Rather than waiting for a conversation to end, we are creating prediction models that estimate customer satisfaction in real-time.
We utilize technical developments in sentiment analysis tools as well as previous observations showing that customer sentiment during a conversation correlates with retrospective customer satisfaction ratings. Considering that real-time customer satisfaction is unknown, we develop unsupervised learning classification models, based on hidden Markov models (HMMs) and sentiment analysis, to categorize the customer state at time t into some arbitrary state space. These states are then mapped to customer satisfaction scores using retrospective prediction models of customer satisfaction.
We study two types of HMMs. In addition to the classical HMM proposed by Baum (1966), we develop a new reactive-HMM that takes into account the agent reactions to customer behavior. We find that the reactive-HMM is more accurate at predicting customer satisfaction in retrospect and hence is recommended for real-time prediction.
This is joint work with Antonio Castellanos, Yair Goldberg, and Galit Yom-Tov.