Evaluating TT-net capability on time series forecasting tasks

Abstract

The Truth-Table rules (TT-rules) [1] framework is a neural network-based rule models that was built upon Truth-Table Net [2]. The method showed promising results in both classification and regression tasks across multiple tabular datasets, and it was expanded to infer reasoning from black-box models such as LLMs [3]. We attempt to further expand this method for time-series forecasting tasks; however, the TT-rules model did not perform up to par for this domain. Thus, We shifted the focus from running LLMs as time series forecasters to exploring different TT-net capabilities on time series forecasting tasks, and investigate why the architecture fails on this specific domain.