Subdomain Identification Strategies for Efficient Machine Learning Models

The performance of machine learning models depends on both the quality of the data and the selection of the models. This work highlights the importance of identifying regions in the input space with similar behavior to improve predictive accuracy. Clustering techniques can partition multivariate time series (MTS) into meaningful subsets, enabling the training of specialized models. We employ k-Medoids and quadtree-based clustering, both using dynamic time warping (DTW) as a similarity measure, and the quadtree also incorporates entropy for partitioning. Long Short-Term Memory (LSTM) networks are trained on these clusters and compared to a global model trained on the entire dataset. The results support the subset modeling hypothesis, showing that models trained in clusters can achieve comparable performance to a global model. This approach offers a comparable alternative that balances prediction accuracy with computational and interpretable advantages.