Weighted sums of random kitchen sinks (W-RKS) are a common feature extraction technique for supervised learning tasks. The method builds features from random projections, which are then weighted and summed to support classification or regression.
A major drawback, as the new arXiv preprint notes, is that many of the extracted random features are redundant or ineffective. This can waste computation and potentially hurt model performance.
The paper, titled "Random Feature Pruning for Multi-Class Classification," directly targets this weakness. Based on the abstract, the work addresses the redundancy problem by pruning random features, with the goal of improving multi-class classification outcomes.
The released abstract is brief and does not include experimental details or specific results, so the exact pruning approach and measured gains are not yet available from the source.