ICE-SEARCH: A Language Model-Driven Feature Selection Approach

• DOI: 10.48550/arxiv.2402.18609
• Publication Date: 2024-02-28
• Authors: Tianze Yang; Tianyi Yang; Shaoshan Liu; Fuyuan Lvu; Xue Liu
• Keywords: Feature (linguistics)

This study unveils the In-Context Evolutionary Search (ICE-SEARCH) method, first work that melds language models (LMs) with evolutionary algorithms for feature selection (FS) tasks and demonstrates its effectiveness in Medical Predictive Analytics (MPA) applications. ICE-SEARCH harnesses crossover mutation capabilities inherent LMs within an framework, significantly improving FS through model's comprehensive world knowledge adaptability to a variety of roles. Our evaluation this methodology spans three crucial MPA tasks: stroke, cardiovascular disease, diabetes, where outperforms traditional methods pinpointing essential features medical achieves State-of-the-Art (SOTA) performance stroke prediction diabetes prediction; Decision-Randomized ranks as SOTA disease prediction. results not only demonstrate efficacy but also underscore versatility, efficiency, scalability integrating tasks. The emphasizes critical role incorporating domain-specific insights, illustrating ICE-SEARCH's robustness, generalizability, swift convergence. opens avenues further research into intricate landscapes, marking significant stride application artificial intelligence predictive analytics.