Reproduce “Post-Hoc Feature Selection Layer for Neural Networks Interpretability” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
João Kenji Suwa, João Pedro Silveira e Silva, Bruno Iochins Grisci
The post-hoc FSL adds a trainable gate before a pretrained tabular model, learning one weight per feature while keeping the model’s parameters fixed. In TabPFN on SynthA, its weighted t-SNE projection separates the two classes more clearly, with silhouette score 0.316 versus 0.186 for Kernel SHAP and 0.070 unweighted. However, Kernel SHAP matches FSL’s PIFS and PSFI scores in this test.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.