Abstract
High-dimensional gene expression data are difficult to analyze with value-based models because absolute expression levels are sensitive to preprocessing and technical variation. Pair-based methods such as TSP and kTSP improve interpretability by using within-sample gene orderings, but they rely on fixed global pair panels.
We investigate whether relational $k$-nearest neighbors (kNN) can provide a useful balance between neighborhood-based flexibility and gene-pair interpretability. Restricted relational representations are constructed from globally informative gene pairs, while individual predictions are explained through compact sets of locally active relations.
The approach is evaluated on seven public gene expression datasets. An ablation study compares Kendall-style and Footrule-based relational variants and is used to select a compact operating point. The selected restricted model is then compared with TSP, kTSP, and full Euclidean and RRM-kNN baselines. The results show the expected trade-off: restricting the relational representation reduces performance relative to full RRM-kNN, but yields compact, sample-specific explanations while remaining competitive with classical pair-based classifiers.
Overall, restricted relational kNN occupies an intermediate position between fixed pair-rule classifiers and full neighborhood-based models, linking global gene-pair selection with local sample-level explanations.
Paper Type
Full Paper
DOI
10.62036/ISD.2026.75
Interpretable Relational kNN for Gene Expression Classification: Compact Global and Local Gene-Pair Explanations
High-dimensional gene expression data are difficult to analyze with value-based models because absolute expression levels are sensitive to preprocessing and technical variation. Pair-based methods such as TSP and kTSP improve interpretability by using within-sample gene orderings, but they rely on fixed global pair panels.
We investigate whether relational $k$-nearest neighbors (kNN) can provide a useful balance between neighborhood-based flexibility and gene-pair interpretability. Restricted relational representations are constructed from globally informative gene pairs, while individual predictions are explained through compact sets of locally active relations.
The approach is evaluated on seven public gene expression datasets. An ablation study compares Kendall-style and Footrule-based relational variants and is used to select a compact operating point. The selected restricted model is then compared with TSP, kTSP, and full Euclidean and RRM-kNN baselines. The results show the expected trade-off: restricting the relational representation reduces performance relative to full RRM-kNN, but yields compact, sample-specific explanations while remaining competitive with classical pair-based classifiers.
Overall, restricted relational kNN occupies an intermediate position between fixed pair-rule classifiers and full neighborhood-based models, linking global gene-pair selection with local sample-level explanations.
Recommended Citation
Kartowicz-Stolarska, I.J. & Czajkowski, M.(2026). Interpretable Relational kNN for Gene Expression Classification: Compact Global and Local Gene-Pair Explanations. In M. Valenta, B. Mannová, R. Pergl, A. Przybylek, M. Lang, H. Linger, C. Schneider, N. Iivari, & E. Insfran (Eds.), Making ISD Sustainable: Reloaded with AI and Automation (ISD2026 Proceedings). Prague, Czech Republic: Czech Technical University in Prague. ISBN: 978-80-01-07585-2. https://doi.org/10.62036/ISD.2026.75