Supervised analysis

Overview

Supervised analysis uses an existing target variable to learn relationships between selected features and an observed outcome.

Ounias preserves the target, feature roles, split configuration, model parameters, predictions, evaluation results, and derived analytical artifacts.

Supported problem types

• Classification — Predict a discrete category or label.
• Binary classification
• Multiclass classification
• Class-level evaluation
• Regression — Predict a continuous numerical outcome.
• Continuous predictions
• Regression-specific evaluation
• Downstream discretization where another analysis requires finite states

Supported supervised methods

• Linear and probabilistic — Logistic Regression: A linear supervised model commonly used as an interpretable classification baseline. [Classification, Linear]
• Linear and probabilistic — Naive Bayes: A probabilistic classifier based on conditional feature distributions. [Classification, Probabilistic]
• Tree based — Random Forest: An ensemble of decision trees capable of capturing nonlinear relationships and interactions. [Classification, Regression, Ensemble]
• Tree based — LightGBM: Gradient-boosted decision trees for higher-capacity supervised modeling. [Classification, Regression, Boosting]
• Tree based — CatBoost: Gradient-boosted decision trees with support for complex nonlinear supervised relationships. [Classification, Regression, Boosting]
• Instance based — K-Nearest Neighbors: Predicts outcomes using nearby observations in the selected feature space. [Classification, Regression, Distance based]

Workflow

• Select the target — Choose the observed outcome the model should learn to predict.
• Define the features — Select the variables available to the model and configure their analytical roles.
• Train and evaluate — Fit the model and evaluate predictions using the configured split.
• Explain and compare — Inspect metrics, errors, explanations, and differences across models.

Evaluation and enrichment

• Held-out performance evaluation
• Standardized metric pillars
• Cross-method comparison
• Feature explanations
• Rule extraction
• Prediction and error inspection
• Stability analysis
• Longitudinal analysis where applicable

Interpretation and limitations

• Predictive performance does not establish causal relationships.
• Evaluation quality depends on the target, split design, dataset, and feature configuration.
• Model metrics should be interpreted together with domain relevance and potential data limitations.