| English | 中文 |
Choose the scientific design using the human-readable examples, then map it to the settings below. Environment setup, configuration-writing code and calls are maintained once in operations.
Start from the complete PM2.5 task template. It expects a prepared station-day table with the following illustrative columns:
| Role | Template field | Adaptation |
|---|---|---|
| Outcome | pm25 |
Actual pollutant statistic, unit and averaging period |
| Key | sample_id |
Unique station-period key |
| Split | dataset |
Existing train, val, test labels from the study design |
| Numeric predictors | temperature, relative_humidity, wind_speed, boundary_layer_height |
Time/space-aligned columns available for the intended use |
| Categorical predictor | season |
Actual category column, declared in categorical_features |
Adapt field names and types, missing markers, coverage rules and split design to the selected data. Perform joins, unit conversions, aggregation and derived-variable construction upstream. The template describes retrospective analysis; forecasting needs an appropriate information-availability definition.
Define the unmonitored spatial/time prediction domain separately from this labelled station table. Complete the input-evidence checks, including ground-observation use in reanalysis/fused/synthetic predictors and overlap with evaluation stations/periods. Record unknown lineage and the chosen independent-reference or sensitivity strategy. Regional inference and field export are not implemented by this template.
The two initial candidates are a depth-5 tree and a 100-tree forest with mtry=2, node size 5. Treat these as a small starting comparison. Set selection_metric to RMSE or MAE and compare additional supported candidates within the four-candidate limit. Check that mtry fits the retained predictor count.
Choose split sizes from the study’s usable data and independent site/time coverage. The template uses the adapter’s supported maximum row caps to avoid inheriting small demonstration sizes; these are ceilings, not requested sample counts or ratios. Review them against the actual cohort and budget before execution. Larger studies need the scaling/sampling decision described in operations.
Review the training median/mode policy and explanation budget, adapt the caps to the intended cohort, and inspect actual selected counts. For screening, choose manual or train_filter and fix the candidate order. For interpretation choose permutation, pdp, ice and optionally shap; use the method’s scientific interpretation rules.
The template has no associated dataset. For a runnable code check use the synthetic parameters; their outcome, predictors and units are synthetic.
Use the complete synthetic PLR task to inspect the interface. A study needs its own justified graph and adjustment set before replacing these fields:
| Decision | Configuration |
|---|---|
| Analysis and outcome | analysis=causal_plr, actual target and target_unit |
| Treatment contrast | causal.treatment, treatment_unit, treatment_scale |
| Adjustment | features=causal.adjustment_set, feature_selection.mode=manual; exclude treatment from predictors |
| Missingness and learner | missing.policy=drop; exactly one nuisance candidate |
| Fold design | causal.folds, resampling.mode, and group_column for group mode |
| Identification | causal.dag and the six justified causal.assumptions |
The branch cross-fits retained training+validation samples. The coefficient uses outcome units per treatment_scale original treatment units. Check graph validity, out-of-fold separation, residual treatment variation and uncertainty in the causal artifacts. Group folds preserve declared groups; independence between groups still belongs to the study design.
Before declaring the assumptions, document the physical quantity or proxy represented by each variable, evidence for causal directions, confounders/mediators, product-generation dependencies and plausible alternative explanations. Plan sensitivity checks for unresolved assumptions. Passing DAG checks or obtaining a precise coefficient does not establish causal reliability.
The existing synthetic causal runner parameters can exercise this branch with the same CLI sequence as prediction. A predictive-importance output is not a substitute for this configuration.
Map a supported numeric impact-prediction task using the same target/feature/split fields. Exposure assignment, lags, outcome denominators and causal identification are established before fitting. For weather normalization, nonlinear exposure-response, dedicated count/binary/survival models, distributed lags or burden calculations, identify the missing operation and implement it in the appropriate skill before composing the workflow.
Use the AirMosaic regression skill.
Question: <prediction or a justified continuous-treatment effect>.
Input: <prepared fixed dataset and dictionary>.
Design: <unobserved prediction domain, labelled reference, target/unit/period, study-specific split>.
Evidence: <physical variable meanings, product lineage, reused observations and issue/release times>.
Causal question, if any: <contrast, justified DAG/adjustment, uncertainty and sensitivity checks>.
Budget: <models, sample caps, seeds and explanation sample sizes>.
Output: <fresh project result destination>.
Read the relevant method, then create personal task/runner JSON using the operations example.
Map each requested operation to a supported setting; identify required adaptations.
Check dependencies and dry-run, execute within the authorized scope, then read artifacts.json.
Check actual cohorts, held-out performance or causal diagnostics and summarize the interpretation.
Reuse skill code for calculations; workflow code supplies settings and orchestration.