Heterogeneous Transfer Learning with Feature Transformation-Based Adaptation for Modeling Dynamical Systems

MingXiao, Zhe Wu

Keywords: nonlinear dynamical system, heterogeneous transfer learning, domain adaptation, generalization error analysis

TL;DR: In this work, a novel heterogeneous transfer learning framework is proposed for modeling dynamical systems, where the source and target domains have different feature spaces.

Abstract: In this work, a novel heterogeneous transfer learning framework is proposed for modeling dynamical systems, where the source and target domains have different feature spaces. A feature transformation scheme is implemented via customized adaptation layers integrated into the pre-trained model. We conduct theoretical analysis of heterogeneous domain adaptation, demonstrating the generalization performance of the pre-trained model on the target domain after feature transformation. Based on this analysis, a two-phase training strategy is proposed to improve the performance of the heterogeneous transfer learning model. The experimental results in four case studies across different application domains demonstrate the effectiveness of the proposed method.

Primary Area: learning on time series and dynamical systems

Submission Number: 12125