AAIB Exam

What is supervised learning?

ML with labeled data; learns input → target mapping for prediction.

Regression vs classification?

Regression predicts continuous values; classification predicts categories/classes.

What is unsupervised learning?

ML with unlabeled data; finds patterns, groups, or hidden structure.

What is reinforcement learning?

Agent takes actions in an environment to maximize cumulative reward.

What is linear regression used for?

Predicting continuous numerical targets.

What is the linear regression objective?

Find weights that minimize prediction error, usually squared loss.

What is regularization?

Penalizing model complexity to reduce overfitting.

What is logistic regression used for?

Binary classification using predicted probabilities.

What does the sigmoid function do?

Maps any real value to a probability between 0 and 1.

How does logistic regression classify?

Uses a probability threshold, often 0.5.

What do logistic regression coefficients represent?

Changes in log-odds.

What is a decision tree?

A sequence of feature-based questions/splits leading to a prediction.

Key decision tree terms?

Root node, split, branch, leaf/terminal node, pruning.

Common decision tree split criteria?

Gini index, entropy, information gain, reduction in variance.

Main advantage and weakness of decision trees?

Advantage: interpretable; weakness: can overfit and generalize poorly.

What is Random Forest?

An ensemble of many decision trees.

What is bagging?

Bootstrapping + aggregating predictions.

How does Random Forest predict?

Classification: majority vote; regression: average prediction.

What is out-of-bag evaluation?

Testing trees on samples not used in their bootstrap sample.

Why is Random Forest better than one tree?

Lower variance, more stability, less overfitting.

What is AdaBoost?

Boosting method that combines weak learners into a strong model.

What weak learner is common in AdaBoost?

Decision stumps: one-split decision trees.

Three core ideas of AdaBoost?

Weak learners, unequal “say,” next learner focuses on previous errors.

How do AdaBoost sample weights work?

Misclassified samples get higher weight in the next round.

What is Gradient Boosting?

Sequentially adds models that reduce previous prediction errors/loss.

Gradient Boosting vs AdaBoost?

AdaBoost reweights samples; Gradient Boosting fits residuals/loss gradients.

What does learning_rate do in boosting?

Controls how much each new tree contributes.

What is LightGBM?

Fast, efficient gradient boosting model for tabular data.

XGBoost vs CatBoost vs LightGBM?

All are boosting models; CatBoost handles categorical variables well, LightGBM is very fast/efficient.

What is K-Means?

Clustering method that assigns points to nearest centroid.

K-Means steps?

Choose k, initialize centroids, assign points, recompute centroids, repeat.

How choose k in K-Means?

Elbow method: choose where improvement starts flattening.

What is DBSCAN?

Density-based clustering that finds dense regions and noise.

DBSCAN point types?

Core, border, noise.

K-Means vs DBSCAN?

K-Means needs k and centroid-like clusters; DBSCAN finds density clusters and noise.

What is feature engineering?

Creating/transformation of variables to improve model learning.

What is data leakage?

When unavailable future/test information enters training.

Why is leakage dangerous?

It gives unrealistically good test results and fails in real use.

What is imbalanced data?

One class is much more frequent than another.

How handle imbalanced data?

Oversampling, undersampling, SMOTE, class weights, threshold tuning.

Why can accuracy mislead on imbalanced data?

Model can predict majority class only and still look accurate.

What is a confusion matrix?

Table of TP, TN, FP, FN.

What is precision?

TP / (TP + FP): how many predicted positives were correct.

What is recall?

TP / (TP + FN): how many real positives were found.

What is F1-score?

Harmonic mean of precision and recall.

Precision vs recall priority?

Precision matters when false positives are costly; recall matters when false negatives are costly.

What is MAE?

Average absolute error.

MSE vs RMSE?

MSE = average squared error; RMSE = square root of MSE in original units.

What is R²?

Proportion of target variance explained by the model.

What is cross-validation?

Repeated train/validation splits for more reliable performance estimation.

Why use TimeSeriesSplit for time series?

It preserves time order and prevents future data leakage.

What is GridSearch?

Tests all predefined hyperparameter combinations.

GridSearch vs RandomizedSearch?

GridSearch tests all combinations; RandomizedSearch samples some, usually faster.

What is AutoML?

Automates model selection, preprocessing, tuning, and evaluation.

What is interpretability?

How understandable the model itself is.

What is explainability?

Explaining why a model made a prediction.

Interpretability vs explainability?

Interpretability = transparent model; explainability = post-hoc reasoning for predictions.

What is SHAP?

XAI method assigning each feature a contribution to a prediction.

SHAP main idea?

Prediction = baseline + feature contributions.

Global vs local explanation?

Global explains overall model behavior; local explains one prediction.

What is trustworthy AI?

AI that is lawful, ethical, and robust.

EU AI Act risk levels?

Unacceptable, high, limited, minimal.

What is Green AI?

AI that values efficiency and computational cost alongside accuracy.

What is deep learning?

ML using neural networks with multiple layers to learn representations.

Core neural network components?

Layers, weights, biases, activation functions, loss function, optimizer.

What is overfitting?

Model learns training noise and performs poorly on new data.

What is an MDP?

Formal RL framework: states, actions, transitions, rewards, discount factor.

Policy vs value vs Q-value?

Policy chooses actions; value rates states; Q-value rates state-action pairs.

Which algorithms were used in the project?

XGBoost, LightGBM, and CatBoost.

Why use boosting models in the project?

They work well on tabular data, capture nonlinear patterns, and handle complex feature interactions.

Why was LightGBM technically strong?

It is fast, memory-efficient, scalable, and performed best across horizons.

What validation method was used in the project?

3-fold TimeSeriesSplit.

Why not use random train-test splitting in the project?

Random splits can leak future information in time-series forecasting.

How does TimeSeriesSplit work?

Train on earlier time periods and validate on later time periods.

What were the main leakage fixes in the project?

Removed the “No” row counter and fixed rolling windows with shift(1).

Why was the “No” column leakage?

It was a sequential row counter correlated with time, not a real predictor.

What was the rolling-window leakage bug?

Rolling means included the current value, which would not be available at prediction time.

Correct rolling feature formula?

PM2.5.shift(1).rolling(window).mean().

Why use shift(1) before rolling?

To ensure only past values are used, not the current/future target.

Main feature types in the project?

Raw pollutant/weather features, lag features, rolling windows, cyclical time, station dummies.

Why use raw numeric features?

Current pollutant and meteorological values are fundamental predictors.

Why use lag features?

Pollution has temporal autocorrelation; recent values predict near-future PM2.5.

Which lag features were used?

1h, 3h, 6h, and 12h lags.

Why use rolling-window features?

They summarize recent average pollution trends.

Why use cyclical time features?

Time is circular, so sin/cos avoids fake jumps like 23 → 0.

Example of cyclical encoding?

hour_sin = sin(2π × hour/24), hour_cos = cos(2π × hour/24).

Why use station dummy variables?

Stations have different spatial/location-specific pollution patterns.

How were missing values handled technically?

Imputation statistics were computed only on training folds and applied to validation/test folds.

Why compute imputation only on training data?

To avoid leakage from future/test data.

How were outliers treated?

Extreme pollution values were flagged but not removed.

Why not remove pollution outliers?

Extreme pollution events are real signals, not necessarily errors.

What scaler was used in the project?

StandardScaler.

Why fit scaling only on training data?

To avoid using test/future statistics during preprocessing.

Do tree-based models require scaling?

Usually no, but scaling can support consistency and comparison.

Best technical project pipeline summary?

Clean data → engineer lag/rolling/time/station features → prevent leakage → TimeSeriesSplit → train XGBoost/LightGBM/CatBoost → evaluate MAE/RMSE.