HowRec

Inspiration

We wanted to explore whether an AI agent could conduct meaningful ML research—not just run a fixed pipeline, but form hypotheses, write code, test ideas, and learn from results.

What it does

HowRec autonomously experiments on a recommender-system ranking task, engineering features, training models, evaluating validation metrics, logging results, and selecting the strongest reproducible submission.

How we built it

We built a guarded research controller around the KuaiRand dataset. It combines train-only feature engineering, model experimentation, validation-based selection, structured experiment logs, failure recovery, and budget/convergence limits.

Challenges we ran into

The hardest parts were preventing data leakage, enforcing reproducible validation, recovering from malformed generated code, and balancing research breadth against limited compute and LLM budgets.

Accomplishments that we're proud of

HowRec discovered that causal popularity features and nonlinear feature interactions improved the validation score over the official FM baseline, while preserving rejected experiments and recovery events for auditability.

What we learned

Autonomous ML research requires more than model training. Clear constraints, reliable evaluation, causal features, experiment tracking, and honest reflection are equally important.

What's next for HowRec

We plan to improve autonomous code generation and recovery, expand research directions, and strengthen the agent’s ability to validate its own hypotheses with less human intervention.

Built With

• feature-engineering
• git
• machine-learning
• numpy
• python