Optimizing prediction models by considering different time granularity of features and target: Problem and solution

Abstract

In many prediction tasks, a common characteristic of training datasets is that features are more frequently updated than target, or in other words, the time granularity, or granularity for short, of features is smaller than that of target. One typical example is predicting ship fuel consumption in maritime transport. Current practice usually ignores such characteristic when developing the prediction models, which may jeopardize prediction accuracy and reliability. However, this issue is neither systematically discussed nor addressed in existing literature. To bridge this gap, this study aims to formally discuss the differences in the granularity of features and target as an ubiquitous issue in prediction problems. Then, an innovative two-stage tree-based approach that considers such differences by maximizing the usage of more frequently updated features is developed. We then go a step further to extend the proposed two-stage tree-based approach to predict accumulative target considering monotonicity and data generation process. Extended numerical experiments using simulated and real datasets in maritime and urban transportation are conducted to verify the superiority of the two-stage tree-based approach and its extension.