Time and context sensitive optimization of machine learning models for sequential data prediction

Abstract

We investigate the nonlinear prediction of sequential time series data through the mixture/combination of machine learning models. First, we introduce a novel ensemble learning approach that effectively combines multiple base learners in a time-aware and context-sensitive manner. This process involves a weight optimization problem targeting a specific loss function while considering (non)convex constraints on the linear combination of base learners. These constraints are theoretically analyzed under known statistics and are automatically incorporated into the meta-learner as part of the optimization process during training. Next, we introduce a direct two-stage approach based on the combination of linear and nonlinear models, where we jointly optimize the parameters of both models to minimize the final regression error. By this joint optimization, we alleviate the well-known underfitting and overfitting problems in modeling sequential data. As the linear model, we use a traditional linear time series forecasting model (SARIMAX) and as the nonlinear model, we use boosted soft decision trees (Soft GBDT). For both of our approaches, we illustrate notable performance improvements on real-life data and well-known competition datasets compared to traditional ensemble/mixture techniques and state-of-the-art forecasting models in the machine learning literature. Additionally, we make the source code of both of our approaches publicly available to facilitate further research and comparison.