Multi-Layer Attention-Based Explainability via Transformers for Tabular Data
Andrea Treviño Gavito, Diego Klabjan, Jean Utke
TL;DR
This work tackles explainability in tabular-data processing by introducing a graph-based, multi-layer attention framework built on transformers. A contextual TD transformer with a priori concept groups is paired with a knowledge-distilled single-head student to preserve multi-layer attention signals for interpretability. Explanations are mined from a DAG constructed from layer-wise attention, using the maximum probability path (and a secondary path) to identify the most impactful feature groups. Across three datasets, MLA is shown to produce stable, concept-group–level explanations and competitive predictions relative to perturbation- and gradient-based baselines, demonstrating practical gains in explainability without sacrificing performance; the approach provides robust, group-level interpretability for TD models.
Abstract
We propose a graph-oriented attention-based explainability method for tabular data. Tasks involving tabular data have been solved mostly using traditional tree-based machine learning models which have the challenges of feature selection and engineering. With that in mind, we consider a transformer architecture for tabular data, which is amenable to explainability, and present a novel way to leverage self-attention mechanism to provide explanations by taking into account the attention matrices of all heads and layers as a whole. The matrices are mapped to a graph structure where groups of features correspond to nodes and attention values to arcs. By finding the maximum probability paths in the graph, we identify groups of features providing larger contributions to explain the model's predictions. To assess the quality of multi-layer attention-based explanations, we compare them with popular attention-, gradient-, and perturbation-based explanability methods.