Relational Feature Modeling: Enhancing Predictive Learning Through Network-Based Features Augmentation

Abstract

Feature representation and augmentation are central factors for improving predictive accuracy of the prediction model, particularly in domains constrained by limited data availability, such as agriculture and biological systems. Data scarcity, usually caused by annotation costs, variability in environmental factors, and the rarity of critical events, restricts the effectiveness of conventional models. While traditional prediction models perform well in data-rich scenarios, their performance degrades in small-data scenarios, leading to reduced predictive accuracy. To address limitations, a network-based feature augmentation prediction framework is proposed for small-data scenarios. An instance-specific parenclitic network is constructed for each data instance by modeling deviations among features, where nodes represent features and edges show the pairwise relationships. Further, three important structural descriptors edge count (global connectivity), entropy centrality (dispersion of influence), and clustering coefficient (local cohesiveness) are extracted and concatenated with the original feature vector to form an augmented representation. The augmented data are then modelled using a regularised multivariate Gaussian distribution, with class-conditional likelihoods. The framework is evaluated on three different datasets and performs better than the baseline methods, with accuracy improvements of 5.1%, 5.6%, and 4.0%. Additional analyses using residual and contour plots confirmed improved stability and interpretability, showing the proposed model is scalable and effective across domains with limited data.