The impact of partial training in Adaptive Piecewise Linear Regression Trees
The 10th International Conference on Future Networks and Distributed Systems, İzmir, Türkiye, 5 - 06 Kasım 2026, ss.1-9, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Basıldığı Şehir: İzmir
- Basıldığı Ülke: Türkiye
- Sayfa Sayıları: ss.1-9
- İstanbul Ticaret Üniversitesi Adresli: Evet
Özet
Piecewise regression trees provide an interpretable approach to modelling non-linear relationships by partitioning the feature space and fitting local predictive models within individual regions. However , when new observations become available, rebuilding the complete tree may be unnecessary if the resulting changes affect only a limited number of regions. This study investigates whether a piecewise Lasso regression tree can instead be adapted through targeted partial retraining while retaining full retraining when broader structural changes are required. A human-in-the-loop framework is introduced in which expert-validated observations provide feedback on model behaviour and identify regions requiring further adaptation. The results show that partial retraining can achieve pre-dictive performance comparable to full retraining while preserving greater stability across successive updates in several experimental settings. Full retraining, however, remains important when new observations indicate changes that cannot be adequately represented by the existing tree structure. These findings demonstrate the potential of selective model adaptation as an alternative to repeatedly rebuilding the complete piecewise regression tree.