Query-Adaptive Local Regression for High-Dimensional Spectral Analysis

This paper proposes Query-Adaptive Local Regression (QALR), a query-specific regression framework for high-dimensional spectral analysis where the number of variables exceeds the number of labeled training samples (p > n). The method replaces global feature representations with instance-specific feature selection through Search Space Reduction (SSR), followed by similarity-based local regression using Softmax-weighted nearest neighbors. The fully vectorized implementation supports efficient CPU, GPU, and TPU acceleration while reducing inference cost. Experimental results on five public spectral datasets demonstrate lower prediction error than conventional chemometric and deep learning methods, together with substantial inference speed improvements for practical real-time spectral analysis.