Glossary¶

This glossary defines the core concepts used across mcgill-showcases.

For each term: - Definition: what it is. - Why it matters: why students and contributors should care. - Where to see it: one or more in-repo projects/artifacts.

Data Profiling And Quality¶

yData Profiling¶

Definition: An automated EDA report that summarizes schema, missingness, distributions, and simple correlations.
Why it matters: It gives a fast first-pass risk scan before modeling.
Where to see it: projects/eda-leakage-profiling-showcase, projects/credit-risk-classification-capstone-showcase.

Univariate Analysis¶

Definition: Per-feature analysis of distribution, central tendency, spread, and missingness.
Why it matters: It detects outliers, constant columns, and suspicious value ranges early.
Where to see it: artifacts/eda/univariate_summary.csv in supervised showcases.

Bivariate Analysis¶

Definition: Relationship analysis between one feature and the target (correlation, target mean by category, etc.).
Why it matters: It exposes predictive signal and potential proxy leakage.
Where to see it: artifacts/eda/bivariate_vs_target.csv.

Missingness Matrix / Missingness Summary¶

Definition: Visualization or tabular report of where and how often values are missing.
Why it matters: Missingness can encode process bias or break model assumptions.
Where to see it: artifacts/eda/missingness_summary.csv, optional missingness_matrix.png.

Feature Type Inference¶

Definition: Classification of columns into logical types such as numeric, categorical, and datetime.
Why it matters: Correct preprocessing choices depend on accurate feature typing.
Where to see it: artifacts/diagnostics/feature_type_summary.csv.

Data Leakage¶

Definition: Information in training features that would not be available at prediction time.
Why it matters: Leakage inflates offline metrics and causes production failure.
Where to see it: artifacts/leakage/leakage_report.csv.

Distribution Shift¶

Definition: A change in data distribution between training and production contexts.
Why it matters: Shift degrades model reliability over time.
Where to see it: projects/mlops-drift-production-showcase.

Covariate Shift¶

Definition: Shift in feature distribution while the target mechanism remains comparatively stable.
Why it matters: Model recalibration or retraining may be needed even if label behavior changes slowly.
Where to see it: drift artifacts in projects/mlops-drift-production-showcase.

Concept Drift¶

Definition: Change in the relationship between features and target over time.
Why it matters: Historic patterns become less predictive, requiring model updates.
Where to see it: retraining decision flow in projects/mlops-drift-production-showcase.

Imputation¶

Definition: Filling missing values using rules such as median, mean, or most-frequent category.
Why it matters: Models and encoders generally require complete matrices.
Where to see it: projects/feature-engineering-dimred-showcase.

Splitting And Validation¶

Train/Validation/Test Split¶

Definition: Three-way data partition: training for fitting, validation for model/threshold selection, test for final unbiased evaluation.
Why it matters: It separates tuning decisions from final reporting.
Where to see it: split manifests across supervised and forecasting showcases.

Stratified Split¶

Definition: Split preserving target class proportions across train/val/test sets.
Why it matters: It stabilizes metric estimates for imbalanced classification.
Where to see it: projects/sota-supervised-learning-showcase, projects/credit-risk-classification-capstone-showcase.

Group Split¶

Definition: Split that keeps related records (same group/query/entity) in a single partition.
Why it matters: Prevents leakage across related entities and avoids optimistic scores.
Where to see it: projects/learning-to-rank-foundations-showcase.

Time-Ordered Split¶

Definition: Split based on chronology (past -> present -> future).
Why it matters: Prevents future information leakage into training.
Where to see it: projects/nyc-demand-forecasting-foundations-showcase.

Cross-Validation (CV)¶

Definition: Repeated train/validation procedures across folds to estimate performance variability.
Why it matters: Reduces reliance on one lucky split.
Where to see it: artifacts/splits/cv_split_manifest.json in split-focused workflows.

Stratified K-Fold¶

Definition: K-fold CV preserving class balance in each fold.
Why it matters: Improves comparability under class imbalance.
Where to see it: split utilities in shared/python/ml_core/splits.py.

No-Overlap Check¶

Definition: Validation that train, validation, and test partitions do not share row identity/index.
Why it matters: Overlap invalidates reported metrics.
Where to see it: no_overlap_checks_passed in split manifests.

Feature Engineering And Representation¶

One-Hot Encoding¶

Definition: Categorical encoding creating one binary feature per category.
Why it matters: Works well for linear/tree models with manageable cardinality.
Where to see it: projects/feature-engineering-dimred-showcase.

Label/Ordinal Encoding¶

Definition: Mapping categories to integer IDs.
Why it matters: Useful for tree models or ordered categories, but risky for linear models when order is artificial.
Where to see it: feature preprocessing in feature engineering and credit risk workflows.

Entity Embeddings¶

Definition: Dense learned vectors representing high-cardinality categorical entities.
Why it matters: Can capture latent similarity better than sparse one-hot vectors.
Where to see it: advanced artifacts in projects/feature-engineering-dimred-showcase.

FeatureTools¶

Definition: Automated feature engineering library for relational/deep feature synthesis.
Why it matters: Speeds up candidate feature generation for tabular tasks.
Where to see it: advanced status outputs in feature engineering showcase.

tsfresh¶

Definition: Time-series feature extraction library creating statistical descriptors from sequences.
Why it matters: Adds rich temporal descriptors for forecasting/classification pipelines.
Where to see it: advanced status outputs in feature engineering showcase.

autofeat¶

Definition: Automated nonlinear feature construction and selection tooling.
Why it matters: Finds transformed features that simple manual pipelines may miss.
Where to see it: advanced status outputs in feature engineering showcase.

RFECV¶

Definition: Recursive feature elimination with cross-validation for subset selection.
Why it matters: Reduces dimensionality while protecting validation quality.
Where to see it: selection outputs in feature engineering showcase.

PCA¶

Definition: Linear projection into principal components maximizing explained variance.
Why it matters: Compresses correlated numeric features and supports visualization.
Where to see it: dimensionality reduction outputs in feature engineering showcase.

t-SNE¶

Definition: Nonlinear embedding optimized for local neighborhood preservation in 2D/3D.
Why it matters: Useful for visual cluster inspection, not primary supervised features.
Where to see it: dimensionality reduction comparisons in feature engineering showcase.

UMAP¶

Definition: Nonlinear manifold projection preserving local and some global structure.
Why it matters: Often faster and more scalable than t-SNE for exploratory visualization.
Where to see it: dimensionality reduction comparisons in feature engineering showcase.

Correlation Matrix¶

Definition: Pairwise correlation table for numeric features.
Why it matters: Highlights redundancy and multicollinearity risk.
Where to see it: artifacts/eda/correlation_matrix.csv.

Imbalanced Learning¶

Class Imbalance¶

Definition: Unequal frequency of target classes (e.g., fraud/non-fraud).
Why it matters: Naive accuracy can look high while minority class performance is poor.
Where to see it: class balance outputs in supervised and credit-risk showcases.

Class Weighting¶

Definition: Increasing loss contribution of minority examples during training.
Why it matters: Often a low-friction baseline for imbalance mitigation.
Where to see it: strategy comparisons in credit-risk and supervised showcases.

Over-Sampling¶

Definition: Increasing minority samples through replication or synthesis.
Why it matters: Helps models see enough minority patterns.
Where to see it: imbalance strategy outputs in supervised/credit-risk workflows.

Under-Sampling¶

Definition: Reducing majority samples to rebalance classes.
Why it matters: Can improve minority recall but may lose information.
Where to see it: imbalance strategy outputs in supervised/credit-risk workflows.

SMOTE¶

Definition: Synthetic Minority Over-sampling Technique creating synthetic minority examples via nearest neighbors.
Why it matters: Provides richer minority coverage than naive duplication.
Where to see it: optional imbalance methods in shared utilities and capstone workflows.

SMOTETomek¶

Definition: SMOTE over-sampling followed by Tomek link cleaning.
Why it matters: Balances classes and removes ambiguous border pairs.
Where to see it: shared/python/ml_core/imbalance.py.

SMOTEENN¶

Definition: SMOTE over-sampling followed by Edited Nearest Neighbors cleaning.
Why it matters: Often improves minority signal quality after synthetic expansion.
Where to see it: shared/python/ml_core/imbalance.py.

Models And Optimization¶

Baseline Model¶

Definition: Simple reference model used before advanced tuning.
Why it matters: Prevents over-engineering and provides a sanity benchmark.
Where to see it: model benchmark outputs in supervised and credit-risk showcases.

LightGBM¶

Definition: Gradient boosting framework optimized for efficiency and strong tabular performance.
Why it matters: Common strong baseline for tabular regression/classification/ranking.
Where to see it: supervised, ranking, and forecasting tracks.

XGBoost¶

Definition: Gradient boosting implementation with robust regularization and mature ecosystem.
Why it matters: Reliable benchmark in many tabular tasks.
Where to see it: classification benchmark outputs in supervised workflows.

CatBoost¶

Definition: Gradient boosting implementation with strong categorical feature handling.
Why it matters: Useful when categorical structure dominates signal.
Where to see it: classification benchmark outputs in supervised workflows.

Stacking¶

Definition: Ensemble technique training a meta-model over predictions of base models.
Why it matters: Can improve robustness when models capture complementary patterns.
Where to see it: ensemble strategy comparisons in supervised workflows.

Hyperparameter Optimization (HPO)¶

Definition: Systematic search over model configuration space.
Why it matters: Reduces manual trial-and-error and improves reproducibility.
Where to see it: projects/automl-hpo-showcase.

Grid Search¶

Definition: Exhaustive evaluation over a fixed hyperparameter grid.
Why it matters: Transparent and reproducible, but expensive at scale.
Where to see it: strategy comparison in AutoML showcase.

Random Search¶

Definition: Random sampling of hyperparameter configurations.
Why it matters: Often more efficient than dense grids in high-dimensional search spaces.
Where to see it: strategy comparison in AutoML showcase.

TPE (Tree-structured Parzen Estimator)¶

Definition: Bayesian optimization method modeling promising vs non-promising regions.
Why it matters: Improves sample efficiency for HPO.
Where to see it: advanced HPO strategy runs in AutoML showcase.

Evaluation And Decisioning¶

ROC-AUC¶

Definition: Area under the ROC curve across classification thresholds.
Why it matters: Measures ranking quality of scores over all thresholds.
Where to see it: supervised and credit-risk evaluation outputs.

PR-AUC¶

Definition: Area under precision-recall curve.
Why it matters: More informative than ROC-AUC on highly imbalanced targets.
Where to see it: supervised and credit-risk evaluation outputs.

Threshold Analysis¶

Definition: Metrics computed across multiple decision thresholds.
Why it matters: Converts model scores into policy-aware decisions.
Where to see it: artifacts/eval/threshold_analysis.csv.

Learning Curve¶

Definition: Performance trend as training data size increases.
Why it matters: Distinguishes data scarcity from model capacity issues.
Where to see it: learning/validation diagnostics in supervised workflows.

RMSE¶

Definition: Root mean squared error; emphasizes larger errors.
Why it matters: Useful when large misses are especially costly.
Where to see it: forecasting and demand API metrics outputs.

MAE¶

Definition: Mean absolute error; average absolute deviation.
Why it matters: Interpretable average error in original units.
Where to see it: forecasting and demand API metrics outputs.

R²¶

Definition: Proportion of variance explained by regression model.
Why it matters: Quick fit-quality indicator, but should not be used alone.
Where to see it: regression-oriented supervised evaluations.

sMAPE¶

Definition: Symmetric mean absolute percentage error for forecast accuracy.
Why it matters: Scale-aware metric common in demand forecasting contexts.
Where to see it: forecasting metrics in NYC demand showcase.

Calibration¶

Definition: Agreement between predicted probabilities and observed outcomes.
Why it matters: Critical for risk scoring and threshold policy reliability.
Where to see it: threshold and probability analysis in classification showcases.

Explainability, Fairness, And Causal ML¶

SHAP¶

Definition: Shapley-value-based feature attribution framework.
Why it matters: Explains local and global model behavior with additive attributions.
Where to see it: projects/xai-fairness-audit-showcase.

LIME¶

Definition: Local surrogate explanation method around individual predictions.
Why it matters: Useful for case-by-case interpretability.
Where to see it: explainability outputs in XAI/fairness showcase.

Fairness Audit¶

Definition: Evaluation of metric disparities across protected or relevant subgroups.
Why it matters: Detects unequal error distribution and policy harm risk.
Where to see it: projects/xai-fairness-audit-showcase.

ATE¶

Definition: Average treatment effect across all units.
Why it matters: Baseline causal estimate for intervention impact.
Where to see it: projects/causalml-kaggle-showcase.

CATE¶

Definition: Conditional average treatment effect for a subgroup/segment.
Why it matters: Supports targeted policy and personalization decisions.
Where to see it: projects/causalml-kaggle-showcase.

Tau (`tau`) / Individual Treatment Effect (ITE)¶

Definition: tau_i = Y_i(1) - Y_i(0) is the unit-level treatment effect for unit i, and tau(x) = E[Y(1) - Y(0) | X = x] is the treatment effect conditioned on features X.
Why it matters: tau connects counterfactual theory to action. Ranking by predicted tau(x) supports treatment policies that target users most likely to benefit.
Where to see it: projects/causalml-kaggle-showcase/notebooks/01_counterfactuals_and_ate.ipynb, projects/causalml-kaggle-showcase/notebooks/02_meta_learners_s_t_x_r.ipynb, projects/causalml-kaggle-showcase/notebooks/04_qini_and_targeting_policy.ipynb.

Examples: - If treated conversion is 0.12 and control conversion is 0.08, the average effect is 0.04 (4 percentage points), which is the sample-level mean of tau.
- If new users have estimated effect 0.06 and returning users 0.01, those are subgroup tau(x) values (CATE-style heterogeneity).
- In uplift ranking workflows, model scores approximate tau(x) and are used to choose who should receive treatment under a budget.

Counterfactual¶

Definition: The unobserved outcome under an alternative treatment/action for the same unit.
Why it matters: Core concept behind causal effect estimation.
Where to see it: causal notebooks and policy simulation outputs.

Uplift Modeling¶

Definition: Modeling incremental outcome caused by treatment versus control.
Why it matters: Improves intervention targeting efficiency.
Where to see it: causal uplift workflows and Qini analysis outputs.

MLOps And Productionization¶

OpenAPI Contract¶

Definition: Machine-readable API schema describing request/response shapes.
Why it matters: Prevents client/server drift and supports contract-first development.
Where to see it: openapi.json in ranking and demand API showcases.

Contract Drift¶

Definition: Mismatch between checked-in OpenAPI schema and runtime API schema.
Why it matters: Breaks client integrations if unchecked.
Where to see it: API openapi-check commands in ranking/demand showcases.

Structured Logging¶

Definition: JSON or key-value logging with consistent fields (trace IDs, route, status).
Why it matters: Enables reliable observability and debugging.
Where to see it: ranking and demand API productization showcases.

RED Metrics¶

Definition: Rate, Errors, Duration service telemetry model.
Why it matters: Core monitoring lens for API reliability.
Where to see it: demand API observability metrics endpoints.

Prometheus Metrics¶

Definition: Pull-based metrics exposed via /metrics in Prometheus format.
Why it matters: Standard, low-friction production telemetry interface.
Where to see it: projects/demand-api-observability-showcase.

OpenTelemetry (OTel)¶

Definition: Standard for traces, metrics, and logs instrumentation across services.
Why it matters: Supports end-to-end tracing and consistent observability semantics.
Where to see it: optional instrumentation hooks in demand API observability showcase.

Canary Rollout¶

Definition: Progressive release where new model/version serves a small traffic subset first.
Why it matters: Limits blast radius of regressions.
Where to see it: projects/model-release-rollout-showcase.

Rollback¶

Definition: Reverting to a previous known-good model/service version.
Why it matters: Essential safety mechanism for production incidents.
Where to see it: rollout decision workflows and registry artifacts.

Model Registry¶

Definition: Versioned catalog of model artifacts and metadata.
Why it matters: Enables traceability, promotion control, and reproducible serving.
Where to see it: production and rollout showcase artifacts.

Forecasting, Ranking, And Policy Terms¶

Horizon¶

Definition: Future time span over which forecasts are generated/evaluated.
Why it matters: Different horizons imply different uncertainty and use cases.
Where to see it: demand forecasting workflows.

Query Group¶

Definition: Set of candidate items ranked together for one ranking request/context.
Why it matters: Ranking loss and metrics are computed per group/query.
Where to see it: projects/learning-to-rank-foundations-showcase.

NDCG¶

Definition: Normalized Discounted Cumulative Gain, a rank-quality metric emphasizing top positions.
Why it matters: Better reflects ranking usefulness than plain accuracy.
Where to see it: ranking metrics artifacts and model evaluation logs.

Regret (Bandits)¶

Definition: Cumulative performance gap between chosen actions and an oracle best action policy.
Why it matters: Measures exploration policy cost over time.
Where to see it: projects/rl-bandits-policy-showcase.

Exploration vs Exploitation¶

Definition: Tradeoff between trying uncertain actions and choosing currently best-known action.
Why it matters: Central decision in online learning and bandit policies.
Where to see it: RL bandits strategy comparisons.