This paper presents a partition-based framework for building and validating regression models through interactive, iterative feature subset selection. The workflow combines quantitative relevance ranking with visual exploration of conditional target distributions. itemize Data Preparation. The dataset is split into training and validation subsets for model fitting and unbiased evaluation. Initial Model Identification. Feature domains are partitioned into disjoint regions, and goodness-of-fit relevance measures are computed for individual features and feature pairs. Ranked small-multiple visualisations of conditional target distributions allow the analyst to identify the most relevant features and characterise relationship structures (non-linearity, local patterns, interactions). The analyst then specifies an initial model (choosing features, transformations, and model type) informed by discovered relationships and domain knowledge. Iterative Model Refinement. The analytical target switches to model residuals. Relevance rankings and conditional-distribution visualisations are recomputed with respect to residuals, exposing effects not yet captured by the model: previously included features rank lower while new relevant features and pair-wise interactions emerge. The analyst discovers additional explanatory variables, transformations, and interactions, and builds a refined model. This cycle of assessment, residual-based exploration, pattern discovery, and model refinement repeats until accuracy is satisfactory. Final Assessment. Model variants are compared through visualisations of local accuracy differences, and the analyst synthesises domain insights about feature–target relationships, interaction effects, and model adequacy.