Cross-domain random forest transfer learning model based on relative entropy regularization
PDF

Keywords

domain adaptation
random forest
relative entropy
feature alignment
transfer learning

Abstract

Domain shift between labeled source data and unlabeled target data often lowers the performance of shallow classifiers such as random forests. This study introduces a cross-domain random forest model that reduces this gap by adding a relative-entropy term to align source and target features and by using a simple weight update to adjust target-domain predictions. The method was tested on the Office-31 dataset under standard transfer pairs. It increased average accuracy by 7.2% compared with TCA-RF and showed more stable results across repeated runs. The improvement was clear on both difficult and easy transfers, indicating that the two added steps help the model adapt to changes in feature distributions without complex training. The findings show that shallow tree ensembles can still perform well in cross-domain settings when supported by simple alignment terms. The study also notes that the method has limits, including the need for enough labeled source data and evaluation only on shared-label tasks.

PDF
Creative Commons License

This work is licensed under a Creative Commons Attribution 4.0 International License.

Copyright (c) 2025 Shira Ben-David, Eitan Rosenberg, Daniel Levi (Author)