New Interpretable Predictive Models Using Proximity Trees and Ensemble Learning

As machine learning models become increasingly complex, the need for interpretability in decision-making processes has grown. This research introduces a novel method for creating inherently interpretable predictive models by identifying and utilizing a small set of representative examples, or 'pivots,' much like how humans make decisions by comparing new situations to past experiences. The proposed approach is built upon a hierarchical, interpretable-by-design pivot selection model that leverages the similarity between these pivots and new input instances. It functions both as a technique for selecting crucial data points and as a standalone predictive model. The innovation extends beyond single pivots to incorporate pairs of pivots, utilizing concepts from proximity and oblique trees, and further enhances its capabilities through ensemble learning. A key advantage of this method is its data modality-agnostic nature, allowing it to process various data types—including tabular data, text, images, and time series—by transforming them using pre-trained networks. Experimental results across these diverse datasets demonstrate that the approach outperforms other instance selection strategies and achieves competitive performance against state-of-the-art interpretable models, all while maintaining a minimal number of pivots for clarity.