Projects

Regret Analysis of the Upper Confidence Bound Algorithm for Multi-Armed Bandits

November 2025
Objective: Authored a self-contained expository paper analyzing the exploration-exploitation tradeoff in multi-armed bandits.

Paper Demo

  • Presented formal definitions of the MAB framework including policies, regret metrics, and the Explore-Then-Commit baseline; derived UCB1’s confidence radius from the Chernoff-Hoeffding bound.
  • Provided a complete proof of UCB1’s O(ln n) regret upper bound, demonstrating how optimism under uncertainty limits suboptimal arm pulls.
  • Built an interactive web demo allowing users to manually explore the bandit problem and compare ETC vs UCB1 policies through Monte Carlo simulations with real-time regret visualization.

Nasdaq Index Prediction and Investment Simulation

March 2025
Objective: Predict future Nasdaq Index using machine learning and neural network models.

View Notebook

  • Preprocessed the dataset (1984 samples, 84 features) including technical indicators, futures contracts, etc. through feature augmentation, conversion, null value handling, and time-series train-test split.
  • Trained four machine learning models (Linear Regression, Decision Tree, Random Forest, XGBoost) and four deep learning models (MLP, LSTM, 1D CNN, CNN-LSTM), with Grid Search for hyperparameter tuning.
  • Implemented Recursive Feature Elimination and Time-Series Cross-Validation for feature selection.
  • Evaluated and compared models and implemented an investment simulation with these models.

CNN and Vision Transformer Pre-trained Models for Facial Emotion Recognition

May 2024
Objective: Build CNN and Vision Transformer networks to recognize facial emotions from scratch.

View Report View Slides

  • Preprocessed the FER-2013 dataset (35,887 grayscale images, 48x48 pixels, 7 categories). Addressed class imbalance with resampling and used data augmentation (rotation, resizing).
  • Designed and optimized CNN and Vision Transformer architectures. Applied learning rate schedules, gradual unfreezing, max pooling (CNN), batch normalization, dropout, and weight decay to address under/overfitting.
  • Achieved accuracy scores of 0.62 (CNN) and 0.66 (ViT). The “happy” category had the highest accuracy (0.83). Models generalized well on images outside the dataset.

Market Data Analysis for Airbnb in Dublin

April 2024
Objective: Understand user profiles and optimize the search-to-booking conversion rate.
Tools: Tableau, Python (Pandas, NumPy, Sklearn, XGBoost, Matplotlib, Seaborn)

View Notebook

  • Preprocessed the dataset (2 tables, 43k rows, 24 columns): handled nulls/outliers, converted data types, merged tables, engineered features.
  • Visualized demographics and booking behavior using Plotly, GeoPandas, and Matplotlib.
  • Trained predictive models (including XGBoost), combined them with a voting classifier, and applied cross-validation. Achieved accuracy of 0.82 (228% higher than random guessing).
  • Suggested business strategies with higher potential ROI.

Analysis on Airline Dataset

September 2023
Objective: Analyze flight data to understand passenger demographics and predict flight status.
Tools: Python (Pandas, Sklearn, Altair)
- Preprocessed dataset (98k rows, 15 columns) using pandas; encoded categorical data with label encoding.
- Visualized passenger profiles (nationality, age, etc.) using Altair.
- Predicted flight status (“on time,” “delayed,” “canceled”) with logistic regression and random forest models.