Data Lab

Experiment

February 2026

Credit card fraud detection

Exploration of an extremely imbalanced transactions dataset, comparing resampling techniques and classification models to detect fraud.

• Six resampling techniques: SMOTE, ADASYN, over-sampling, under-sampling, SMOTETomek and SMOTEENN.
• Three models compared: logistic regression, random forest and gradient boosting.
• Imbalance-aware evaluation: precision, recall, F1, ROC-AUC, average precision and precision-recall curve, with a plain-language interpretation.

LimitThe sample shipped in the repository is too small to draw conclusions: this is an educational project, not a production fraud system.

Analysis

January 2026

Census income prediction

Predicting whether income exceeds 50,000 USD: data exploration, training and comparison of six models, then prediction from a form.

• Six models compared: k-nearest neighbours, decision tree, random forest, gradient boosting, logistic regression and SVM.
• Standardised features, configurable test size and random seed, confusion matrix of the best model.
• The best model is saved then reloaded for prediction, with a fallback model if the file is missing.

LimitOnly numeric features are used; categorical variables are not yet included.

Experiment

May 2026

Student depression risk score

A Streamlit app that estimates a 0 to 100 % risk score from lifestyle and academic factors, using a random forest.

• Random forest in a scikit-learn pipeline, with documented hyperparameters.
• Three-level risk gauge and a radar chart comparing a profile with dataset averages.
• Pinned dependencies so the saved model reloads without version conflicts.

LimitMachine Learning course project trained on survey data: it is not a medical tool.