Simultaneous Co-segmentation and Predictive Modeling for Large, Temporal Marketing Data

Several marketing problems involve prediction of customer purchase behavior and forecasting future preferences. We consider predictive modeling of large scale, bi-modal or multimodal temporal marketing data, for instance, datasets consisting of customer spending behavior over time. Such datasets are characterized by variability in purchase patterns across different customer subgroups and shifting trends in behavior over time, which pose challenges to any predictive technique. The response variable in this case can be viewed as the entries of a matrix/tensor, while the independent variables are the attributes associated with different modes. We propose a simultaneous co-segmentation and learning approach that partitions the input space into relatively homogeneous regions by simultaneously clustering the"customers", segmenting the "time" axis and concurrently learning predictive models for each (evolving) homogeneous partition. This approach forms a very general framework for predicting missing entries in the data matrix/tensor as well as for making predictions for new entities, e.g., new customers or future time intervals. We illustrate the effectiveness of our approach through detailed experimentation on the challenging ERIM marketing dataset.