Learning Predictive, Online Approximations of Explanatory, Offline Algorithms

Mattson Thieme, Ammar Gilani, Han Liu

Keywords: multi-task learning, machine learning, online algorithms, offline algorithms

Abstract: In this work, we introduce a general methodology for approximating offline algorithms in online settings. By encoding the behavior of offline algorithms in graphs, we train a multi-task learning model to simultaneously detect behavioral structures which have already occurred and predict those that may come next. We demonstrate the methodology on both synthetic data and historical stock market data, where the contrast between explanation and prediction is particularly stark. Taken together, our work represents the first general and end-to-end differentiable approach for generating online approximations of offline algorithms.

One-sentence Summary: We introduce a novel, multi-task learning methodology to approximate the behavior of arbitrary offline algorithms in online settings.