Evaluating the Added Value of Pictorial Data for Customer Churn Prediction

The purpose of this paper is to evaluate whether pictorial data can improve customer churn prediction and, if so, which pictures are most important. We use Random Forest and five times twofold cross-validation to analyze how much pictorial stimulus-choice data increase the AUC and top decile lift of a churn model over and above administrative, operational, complaints and traditional survey data. We found that pictorial-stimulus choice data significantly increase models with administrative and operational data. The most important pictures are facial expressions, colors, and motivational scenes. Pictorial variables can reach importance of up to 21% of the importance of the best predictor included in the predictor set. Given this importance, managers could mine pictorial data from social media sites (e.g., Pinterest.com) in order to augment their internal customer database. To the best of our knowledge this study is the first that evaluates the added value of pictorial stimulus-choice data in predictive models. This is important because social media platforms are increasingly sharing their data and because of the recent rise of social media based on pictures. Pictorial data may soon become a viable option for data data-augmentation.