Data Reduction Methods for Regression: A Multi-objective Perspective

Data reduction plays a central role in predictive data analysis, particularly in regression tasks, where increasing data volume and noise levels can adversely affect three key objectives: predictive performance, model interpretability, and computational efficiency. Selecting appropriate data reduction methods in such settings requires balancing these competing objectives, yet systematic and reproducible support for making such choices remains limited. This paper suggests a context-aware, multi-objective framework for data reduction method selection for regression analysis. Given a dataset characterized by its technical properties and a user-defined priority ordering over the considered objectives, the framework empirically characterizes candidate reduction methods, identifies Pareto-efficient trade-offs, and produces ranked recommendations using a lexicographic selection strategy consistent with the specified priorities. The framework is evaluated on real-world regression datasets with publicly available analytical pipelines containing explicitly implemented data reduction choices. The results show that the proposed approach frequently reproduces practitioner-selected reduction strategies under comparable priorities, while also exposing meaningful alternative trade-offs when different objectives are emphasized. These findings indicate that context-aware, preference-aware selection can provide effective and transparent support for data reduction decisions in regression-oriented analytical pipelines.