Skip to content

User Guides

Task-oriented guides, one per topic. Each one explains the task and shows runnable examples of the key APIs. If you have not run SMF before, start with the Quickstart.

Guide List

  • Sample Management

    Split, stratify, and balance a modeling sample, and search for a good split seed. SampleSplitter, StratifiedSampler, SampleBalancer, select_sample_seed

  • ODPS Data Extraction

    Run SQL on Alibaba Cloud MaxCompute, and download and upload tables. ODPSRunner, parse_sql_file, pull_attributes_in_batch

  • Data Consistency Comparison

    Compare two DataFrames or large CSV tables, in the spirit of SAS proc compare. ProcCompareEngine, proc_compare

  • Pipeline GUI Schema

    Read pipeline config metadata to build a configuration form, and export or validate configs as YAML. extract_pipeline_schema, config_to_yaml, validate_pipeline_config

  • WOE Encoding

    Bin variables and encode them with weight of evidence. WOE_Master, MonotoneWOEBinner, is_monotonic

  • WOE Binning Engine

    Reuse one fitted binning engine for PSI, IV, correlation screening, and the final model. as_woe_engine, binning_engine, woe_binner

  • Feature Screening

    Screen features by stability, information value, and correlation. PSICalculator, VarExtractionInsights, CorrelationFilter, feature_screen

  • Model Training

    Train logistic regression and gradient-boosting models, optionally with sample weights, and eliminate variables. LRMaster, GradientBoostingModel, BackwardVariableEliminator

  • GBM Hyperparameter Search

    Search LightGBM, XGBoost, or CatBoost parameters with a grid or Optuna, judged on INS, OOS, and OOT datasets. GradientBoostingModel.param_search

  • Model Registry and Versioning

    Save a model with metadata (version, features, WOE mapping path, metrics) and load it back. save_model, load_model, load_model_metadata

  • Model Evaluation

    KS, AUC, Gains tables, score comparison, and slice-by-slice evaluation, weighted or unweighted. PerformanceEvaluator, GainsTableCalculator, Model_Evaluation_Tool, EvaluationPipeline

  • Model Explainability

    Explain a model with SHAP, Owen value, PDP, ICE, ALE, and LIME. ModelExplainer

  • Reject Inference and Distribution Adaptation

    Correct selection bias in a sample of approved applicants. RejectInferenceFactory, DistributionAdaptation

  • Online/Offline Consistency Check

    Check that online and offline scores and features agree before go-live. UATConsistencyChecker, UATConfig

  • Excel Report Generation

    Write formatted Excel workbooks from modeling results. ExcelMaster, Template, Report

Release-by-release changes are collected under the ChangeLog tab.

Suggested Reading Order

  1. Run the Quickstart end to end.
  2. Read End-to-End Modeling Pipeline for the full manual workflow, then the guides in workflow order: samples, WOE, feature screening, models, evaluation, explainability, and Excel reports.
  3. If you want PSI, IV, correlation screening, and the final model to share one binning, read WOE Encoding and then WOE Binning Engine.
  4. To run the whole workflow in one call, see Top-Level Pipelines.
  5. To look up a signature, go to the API Reference.