A team of researchers has built a multi-agent AI system that designs its own feature-selection algorithms instead of relying on a human expert to hand-tune one for each dataset.
The system, called FeatureHospital, targets multi-view multi-label feature selection, the process of picking which data attributes matter most when a dataset combines multiple sources of information and multiple labels per example. It starts by diagnosing a dataset to flag its specific feature-selection issues. Specialist agents, each equipped with a domain 'Skill,' then prescribe optimization strategies and loss terms suited to those issues. The prescriptions are reconciled to strip out overlaps and resolve conflicts, then merged into one dataset-specific objective that gets optimized to produce the final feature subset.
Feature-selection algorithms are usually hand-built, tailored to a dataset's particular quirks, which takes expert knowledge and real time. FeatureHospital automates that judgment call by splitting it into diagnosis, prescription, and reconciliation steps handled by separate agents, echoing how AutoML tools automate model tuning but applying the idea to a less-automated corner of machine learning. If the approach holds up outside the paper, it could cut the manual tuning that currently keeps good feature selection out of reach for teams without a dedicated specialist.
The evidence so far comes from the researchers' own experiments, not independent replication, so this reads as a promising first draft of an automated feature engineer, not a proven replacement for one.