Mixing Configurations for Downstream Prediction

Clustering-based features are widely used in machine learning, but most methods must choose a resolution---a choice that is global, fixed, and ad hoc. Recent work shows that varying the resolution parameter produces only a finite set of structurally stable partitions, known as configurations. Based on this, we introduce Configuration-Mixed Prediction (CMP), a setting where models learn to adaptively weight these configurations per sample for downstream prediction. We propose MixConfig, a plug-and-play feature augmentation module that extracts configurations from any embedding and learns energy-aware mixing weights via a novel selector that jointly reasons about sample context, cluster assignments, and stability statistics. Experiments across tabular, molecular, vision, and text domains demonstrate consistent improvements over single-resolution and static baselines across diverse predictor architectures, with gains particularly pronounced in low-data regimes.

Many real-world data sets can be understood at different levels of detail. For example, an image may be grouped by broad object type or by fine visual details, and a molecule may be described by large chemical groups or smaller bond patterns. Most machine learning methods that use such group information must choose one fixed level of detail, even though different examples may need different levels. This paper introduces MixConfig, a method that lets a prediction model use several useful levels of grouping at once. Instead of forcing the model to rely on one chosen grouping, MixConfig finds a small set of stable groupings in the data and learns how much each grouping should matter for each individual example. The method can be added to existing prediction systems without redesigning their main architecture. Across tasks involving tables, images, molecules, and text, MixConfig improves prediction performance over methods that use either no grouping information or only one fixed grouping. The gains are especially clear when labeled data is limited, suggesting that multi-level structure can help models learn more from fewer labels.