The Least Absolute Shrinkage and Selection Operator (Lasso) is widely used for feature selection in linear models and has been extended to multi‑layer perceptrons (MLP) for both individual and group feature selection. Group‑Lasso has similarly been applied to feed‑forward neural networks. This work is the first to embed the Group‑Lasso principle into a Siamese network architecture, aiming to select a compact feature subset that preserves the intrinsic representation of the input while achieving high classification accuracy. Two variants are proposed: one based on Euclidean distance and another using a modified Jaccard similarity tailored for real‑valued vectors, the latter targeting high‑dimensional sparse data. Experiments on twelve benchmark datasets (feature dimensions ranging from 6 to 7070) with eight classifiers demonstrate that the Siamese Lasso consistently identifies compact, discriminative feature subsets that maintain competitive or superior classification performance, even though the loss function does not explicitly include a classification regularizer. The quality of the learned embeddings is further examined with manifold learning techniques such as UMAP and t‑SNE.
Review