Agent Auto Sci AI ML

Use this subskill when the project needs modeling rather than only descriptive analysis.

Fast Workflow

1. Define prediction or explanation target.
2. Build a data dictionary and leakage checklist.
3. Establish simple baselines before complex models.
4. Choose model family: classical ML, deep learning, time series, spatial ML, or NLP/transformer.
5. Split data correctly: spatial, temporal, group, or external validation as needed.
6. Evaluate with task-appropriate metrics and uncertainty.
7. Interpret with SHAP, permutation importance, partial dependence, or ablation.
8. Translate model results into manuscript claims only after robustness checks.

Read references/ml_ai_workflows.md.

For full paper projects, this skill owns baseline design, model/method selection, leakage checks, train/test split logic, hyperparameter search, ablation, robustness, error-case analysis, XAI, and methods/results reporting boundaries. It must work with agent-auto-sci-methodology before turning feature importance into mechanism language.

For deeper K-Dense-style encapsulation:

• references/k_dense_ml_ai_mapping.md: how scikit-learn, PyTorch Lightning, transformers, SHAP, time-series, and XAI skills are adapted.
• references/sport_geography_ml_playbook.md: sport geography modeling patterns, leakage checks, baselines, and manuscript reporting.

Related Helper Skills

Use installed helper skills when useful:

• scikit-learn
• pytorch-lightning
• transformers
• shap
• aeon
• timesfm-forecasting
• statistical-analysis
• agent-auto-sci-data-viz
• agent-auto-sci-geospatial

Must Not Do

• Do not use ML if the research question only needs transparent spatial/statistical analysis.
• Do not report accuracy without baselines and uncertainty.
• Do not mix train/test spatial units when spatial leakage is possible.
• Do not use SHAP as causal evidence.