Skip to main content

Machine Learning & Choice Models

Statios provides machine learning and choice modeling tools: supervised learning with explainability (Random Forest, Gradient Boosting, SVM, Decision Trees, neural networks); model evaluation (cross-validation, ROC/AUC, confusion matrix, feature importance, SHAP); and conjoint analysis (traditional, choice-based, hierarchical Bayes) for preference modeling.

31 procedures · 8 with documented limitations. Every result is computed by the open Python engine and is exportable to APA tables and reports.

ProcedureWhat it doesStatusKey reference
Random ForestFit random forest for classification or regression with feature importances and OOB score.✅ FullBreiman (2001)
Gradient BoostingFit gradient boosted model (XGBoost if available) with learning curves and feature importances.✅ FullPedregosa et al. (2011)
Support Vector MachineFit SVM for classification or regression with standardized features and support vector counts.✅ FullPedregosa et al. (2011)
Decision TreeFit decision tree with tree rules, feature importances, and depth/leaf metrics.✅ FullPedregosa et al. (2011)
Cross-ValidationK-fold cross-validation with train/test scores and timing metrics across folds.✅ FullPedregosa et al. (2011)
ROC/AUC AnalysisBinary classification ROC curve with AUC and Youden's J optimal threshold.✅ FullPedregosa et al. (2011)
Confusion MatrixMulti-class confusion matrix with per-class precision/recall/F1, normalizations, and macro/weighted averages.✅ FullPedregosa et al. (2011)
Classification MetricsAccuracy, precision, recall, F1, Matthews correlation, Cohen's kappa, and log-loss.✅ FullPedregosa et al. (2011)
Feature ImportanceBuilt-in and permutation-based feature importances with standard errors.✅ FullPedregosa et al. (2011)
SHAP ValuesSHAP values via TreeExplainer (exact TreeSHAP) or KernelExplainer (model-agnostic) with expected values.⚠️ LimitedPedregosa et al. (2011)
SHAP Summary PlotSHAP summary (beeswarm) plot data with feature importance ranking and per-sample contributions.✅ FullPedregosa et al. (2011)
Train/Test PredictFit model on train set and return test predictions, actuals, residuals/probabilities.✅ FullPedregosa et al. (2011)
MLP ClassifierScikit-learn multilayer perceptron for classification with configurable hidden layers and activation.✅ FullPedregosa et al. (2011)
MLP RegressorScikit-learn multilayer perceptron for regression with configurable hidden layers and activation.✅ FullPedregosa et al. (2011)
Deep Neural NetworkKeras/TensorFlow deep network for classification or regression with custom architecture, epochs, batch size.✅ FullPedregosa et al. (2011)
AutoencoderUnsupervised autoencoder for dimensionality reduction and feature learning.✅ FullPedregosa et al. (2011)
CNN for Tabular DataConvolutional neural network adapted for tabular classification or regression.✅ FullPedregosa et al. (2011)
Neural Network Hyperparameter TuningGrid search over neural network hyperparameters with cross-validation.✅ FullPedregosa et al. (2011)
CHAIDChi-squared automatic interaction detection for tree-based classification.⚠️ LimitedPedregosa et al. (2011)
QUESTQuick, unbiased, efficient statistical tree method for classification.⚠️ LimitedPedregosa et al. (2011)
C5.0C5.0 decision tree and rule learner for classification.⚠️ LimitedPedregosa et al. (2011)
Ensemble Trees ComparisonSide-by-side comparison of multiple tree-based ensemble methods.⚠️ LimitedPedregosa et al. (2011)
Tree ExportExport trained tree model in formats suitable for deployment.⚠️ LimitedPedregosa et al. (2011)
Compare TreesComparative metrics across tree models (accuracy, depth, interpretability).⚠️ LimitedPedregosa et al. (2011)
Traditional Conjoint AnalysisRating-based conjoint via OLS with dummy-coded attributes and part-worths.✅ FullSeabold & Perktold (2010)
Choice-Based ConjointAggregate multinomial logit conjoint from choice data with part-worths and hit rate.✅ FullSeabold & Perktold (2010)
Hierarchical Bayes CBCEmpirical-Bayes iterated-MAP CBC with individual part-worths (not true HB-MCMC).⚠️ LimitedSeabold & Perktold (2010)
Market SimulatorMarket share simulation given part-worths and scenarios (logit, first-choice, randomized).✅ FullSeabold & Perktold (2010)
Attribute ImportanceRelative attribute importance from part-worth ranges (% allocation).✅ FullSeabold & Perktold (2010)
Conjoint DesignFractional experimental design generation (full, orthogonal, D-optimal).✅ FullSeabold & Perktold (2010)
Conjoint SegmentationSegment respondents using K-means on individual part-worths.✅ FullSeabold & Perktold (2010)

Notes & limitations

  • SHAP Values — TreeExplainer unavailable in browser (WASM); fallback to KernelExplainer (sampled Shapley values, not exact TreeSHAP).
  • CHAID — Stub implementation; returns placeholder.
  • QUEST — Stub implementation; returns placeholder.
  • C5.0 — Stub implementation; returns placeholder.
  • Ensemble Trees Comparison — Stub implementation; returns placeholder.
  • Tree Export — Stub implementation; returns placeholder.
  • Compare Trees — Stub implementation; returns placeholder.
  • Hierarchical Bayes CBC — Empirical-Bayes / EM-style scheme, NOT Gibbs/Metropolis; samples are posterior modes, not draws; heterogeneity SD understates posterior uncertainty.