🤖 AI Summary
This work addresses the fundamental limitation of classical k-nearest neighbors (k-NN) in tabular data modeling—its non-differentiability, which precludes end-to-end optimization. We propose the first deep differentiable Neighborhood Component Analysis (NCA) framework. Methodologically, we integrate NCA directly into deep neural network representation learning, eliminating hand-crafted feature engineering and conventional dimensionality reduction; the framework incorporates stochastic augmentation, an adaptive loss function, and ensemble-based prediction for fully end-to-end training. Contributions: (i) We empirically establish, for the first time, that a pure NCA-based model achieves strong competitive performance on standard tabular benchmarks; (ii) augmented with deep representations and training stochasticity, our method attains state-of-the-art results across 300 classification and regression benchmarks—significantly outperforming leading deep tabular models (e.g., TabNet, SAINT) and matching the performance of advanced gradient-boosted tree methods (e.g., CatBoost). The code is publicly available.
📝 Abstract
The widespread enthusiasm for deep learning has recently expanded into the domain of tabular data. Recognizing that the advancement in deep tabular methods is often inspired by classical methods, e.g., integration of nearest neighbors into neural networks, we investigate whether these classical methods can be revitalized with modern techniques. We revisit a differentiable version of $K$-nearest neighbors (KNN) -- Neighbourhood Components Analysis (NCA) -- originally designed to learn a linear projection to capture semantic similarities between instances, and seek to gradually add modern deep learning techniques on top. Surprisingly, our implementation of NCA using SGD and without dimensionality reduction already achieves decent performance on tabular data, in contrast to the results of using existing toolboxes like scikit-learn. Further equipping NCA with deep representations and additional training stochasticity significantly enhances its capability, being on par with the leading tree-based method CatBoost and outperforming existing deep tabular models in both classification and regression tasks on 300 datasets. We conclude our paper by analyzing the factors behind these improvements, including loss functions, prediction strategies, and deep architectures. The code is available at https://github.com/qile2000/LAMDA-TALENT.