Interpretable Lazy Classification with Interval Pattern Structures and Local Interval Explanations

International Journal of Approximate Reasoning. 2026. Vol. 197. Article 109754.

Tomat A., Sergei O. Kuznetsov

Interval Pattern Structures (IPS) provide a natural way to represent local, human-readable explanations for predictions on numerical data through vectors of intervals interpreted as axis-parallel hyper-rectangles. In this paper, we develop and evaluate an IPS-based k-nearest neighbors classifier, IPS-KNN, that explains each prediction through a single local interval description rather than through the aggregation of many candidate descriptions. The proposed method introduces three explanation mechanisms: the Reason for Classification (RC), the Reduced Reason for Classification (RRC), and local feature importance scores derived from information gain. The method is formulated in the multiclass setting, while its main explanatory output remains a local interval description for each prediction. To contrast single-description and aggregation-based IPS strategies under the same protocol, we also include deterministic and randomized aggregation-based IPS baselines. Experiments on twelve numerical datasets show that IPS-KNN generally outperforms the aggregation-based IPS baselines, remains competitive with standard distance-weighted k-NN and several strong non-interpretable baselines, and produces compact local explanations for individual predictions. These results indicate that IPS-KNN combines competitive predictive performance with compact local interval-based explanations for numerical classification.

Keywords: Formal concept analysis, k-nearest neighbors, Lazy classification, Interpretable machine learning, Interval pattern structures, Local explanations