Project work from the Mastering Applied Data Science with Deep Learning programme, covering exploratory analysis, machine learning and deep learning on real-world datasets. Each project links to its own GitHub repository with the full notebook.
A feed-forward neural network built with TensorFlow/Keras to predict Titanic passenger survival from six standardised features. Training for 100 epochs revealed clear overfitting, with validation accuracy falling from 0.88 to around 0.81. Retraining with early stopping on validation loss kept the best weights and held validation accuracy at about 0.87.
View project on GitHub →Predicts sale prices for 21,613 homes in the Seattle area (2014–2015). After removing extreme outliers and engineering features such as house age, renovation and basement flags, zip-code categories and squared size variables, thirteen regression models are compared. XGBoost meets the project targets with R² 0.85 and RMSE of about $83,000, with Lasso and Ridge close behind.
View project on GitHub →Analyses 10,417 Yelp reviews of Bacchanal Buffet in Las Vegas. Reviews are cleaned, language-filtered and lemmatised, then eight classifiers are compared across four text representations; Logistic Regression with n-grams performs best (accuracy 0.83). VADER and TextBlob scores, word clouds and n-gram counts show that complaints centre on long waits and value for money, while praise centres on the food, above all the crab legs.
View project on GitHub →Predicts the price of 1,303 laptop models from their specifications. Raw text columns are turned into usable features: pixel density from screen size and resolution, separate SSD and HDD capacities, CPU groups and clock speed, GPU brand and touchscreen/IPS flags. Five models are compared on log price; XGBoost performs best (R² 0.90), ahead of a GridSearchCV-tuned Random Forest. RAM and SSD capacity emerge as the strongest price drivers.
View project on GitHub →Groups 138 countries by military capability using K-Means on 28 indicators covering air, land and naval forces, defense spending, manpower and geography. After correcting a data misalignment and treating structural missing values (landlocked and island countries), three profiles emerge: Major Powers, Regional Powers and Limited Capacity. The Global Firepower Power Index, left out of the model, ranks the clusters in the same order, independently validating the result.
View project on GitHub →Segments 3,054 e-commerce customers for targeted marketing. Raw order data is split into customer, product and order tables in a SQLite database, then summarised as Recency, Frequency and Monetary values and clustered with K-Means (elbow, silhouette and business judgement). Four segments emerge: a loyal group of under 3% of customers brings in a fifth of revenue, while lapsed big spenders hold over half of it, the clearest target for a win-back campaign.
View project on GitHub →Multiple linear regression predicting sales from TV, radio and newspaper advertising budgets across 200 markets. The model explains 91% of the variation in sales on held-out test data (RMSE 1.68). Holding other channels constant, radio delivers the largest sales gain per unit of spend, followed by TV, while newspaper spending adds almost nothing.
View project on GitHub →Groups 65 airlines at San Francisco International Airport by traffic profile, using 13 years of monthly passenger records (2005–2018). Raw records are aggregated to one row per airline and described by passenger volume, domestic, low-fare, transit and long-haul shares. K-Means with elbow and silhouette analysis (silhouette 0.70) identifies six profiles, from long-haul international carriers to low-fare and regional domestic airlines.
View project on GitHub →End-to-end credit-risk model on 257,000 loan records: deduplication, detection and removal of two target-leakage sources, ratio features (loan-to-income, debt-to-income, credit utilisation) and a comparison of seven classifiers. Logistic Regression is recommended for transparency (ROC AUC 0.70); at a 0.5 threshold it cuts the default rate among approved loans from 27.4% to 17.6%.
View project on GitHub →Predicts passenger survival on the Kaggle Titanic dataset. Missing ages are imputed with ticket-class medians, categorical variables are encoded, and Random Forest and Gradient Boosting classifiers generate predictions for the 418-passenger test set.
View project on GitHub →