📌 Project Overview: Developed for a Bayesian Optimisation (BO) competition to optimise eight unknown black-box functions across domains such as contamination detection, drug discovery, and hyperparameter tuning. Functions were high-cost to evaluate — queries allowed only every few days — simulating real-world constraints.

🔍 Problem Breakdown

Each function reflected a specific industrial/scientific task:

• Function 1 – 2D: Radiation Source Detection
• Function 2 – 2D: Noisy ML Simulator
• Function 3 – 3D: Drug Discovery
• Function 4 – 4D: Inaccurate Business Model
• Function 5 – 4D: Chemical Yield
• Function 6 – 5D: Cake Recipe Optimisation
• Function 7 – 6D: ML Hyperparameter Tuning
• Function 8 – 8D: High-Dim Black-Box Search

Initial observations were provided in .npy format to seed Gaussian Process (GP) fitting.

📚 Research-Informed Strategy

Inspired by state-of-the-art work:

• HEBO (Huawei, NeurIPS 2020)
• Warped GPs (Snelson & Rasmussen, NIPS 2003)
• Acquisition Blending (arXiv:2206.03301)

These guided our approach to transformations, noise handling, kernel structure, and acquisition strategy blending (EI + PI + UCB via NSGA-II).

🛠️ Tools & Technologies

• Language: Python 3.10+
• Libraries: numpy, scikit-learn, scipy, matplotlib, pymoo
• GP Kernels: RBF, Matern, RationalQuadratic, Constant, White
• Optimisation: NSGA-II (pymoo) for acquisition mixing
• Exploration: ε-greedy, dynamic bound restriction
• Diagnostics: PCA, Random Forest feature importance

⚙️ End-to-End Pipeline

• Problem Understanding: Defined bounds from context and units
• EDA: Outlier checks, noise estimation, distribution plots
• Transformations: Yeo–Johnson, Box–Cox (output), Kumaraswamy or MinMax (input)
• Kernel Design: RBF for smoothness, Matern for nonlinearity, White for noise
• GP Structure: Constant × (RBF + Matern) + White
• Acquisition: EI + PI + UCB (MACE routine)
• Iteration Loop: Fit → Optimise → Submit → Update → Log mean/variance
• Visualisation: PCA to reduce and track GP uncertainty & decisions

🧠 Key Insights

• Reusable pipeline enabled fast, consistent strategy testing
• Dynamic bound shrinking & feature filtering helped in high-D settings
• Warping + WhiteKernel tuning improved generalisation under noise
• Transformation choices heavily affected GP smoothness & acquisition
• NSGA-II acquisition mixing improved late-stage exploration vs exploitation

✅ Results

Delivered strong results across the board, with top performance in Function 6. Applied advanced BO techniques in a realistic, high-cost query environment.

Predict-When-They-Churn

Behaviour-Driven Customer-Loss Forecasting

Online shops lose revenue when loyal buyers quietly disappear. Imagine you want to know who is quietly slipping away before it’s too late. This project watches how long shoppers stay, how often they return, and what they buy to flag customers who are unlikely to purchase again during the next 180 days. By turning millions of raw click-stream records into clear "stay-or-leave" scores, the model can let marketing teams run retention offers before the customer is gone.

📦 Data

• Source: Retail Rocket – 2.7M events
• Files: events.csv, item_properties.csv, category_tree.csv
• Ingestion: Loaded via PySpark with explicit schemas
• Note: If property = 790, then value is price

🤖 Model

• Algorithm: RandomForestClassifier (scikit-learn)
• Why: Best trade-off between recall and interpretability
• Features: 158 per visitor – session stats, recency, ratios
• Output: Churn label (1 = likely churn in 30 days)

Project Overview

💡 Key Features

⚙️ How It Works

🛠️ Technologies Used