Sanjeeb Adhikari

The rapid growth rate of electric cars (EVs) has brought to light serious deficiencies with the geographical distribution of public charging stations that serve EV drivers. Although the Federal Government has allocated a large amount of money to deploy more stations and boost the availability of charging locations, deployment strategies still rely primarily on reactive and supply side considerations. This proposal outlines a data framework for identifying underserved areas within the EV charging network in the U.S. through the use of supervised machine learning analysis. An XGBoost regression model will analyze the factors affecting usage patterns in more established EV markets (California, Colorado, and Vermont) so that it can develop a predictive demand profile based on factors such as demographics, transportation, and built environment (population density, traffic volume, and housing type, etc.). Once the model has been successfully trained with established markets, it will be used to assess latent demand in both emerging and underserved regions of the U.S. By comparing predicted demand against current charging capacity, the study will quantify deficiencies in charging station infrastructure on a U.S. Census Tract basis, and produce a prioritized list of locations where future charging station investment would be most appropriate.

This project develops a news recommendation system using contextual bandit algorithms on the Microsoft News Dataset (MIND). The objective is to recommend relevant news articles by learning from user interactions (click and no-click) while effectively balancing exploration and exploitation. Multiple algorithms, including Epsilon-Greedy, UCB, Thompson Sampling, LinUCB, and Disjoint LinUCB, were implemented and compared. Two versions of the system were built: a basic model using user history and news embeddings, and an enhanced model that incorporates additional features such as entity embeddings and date-time information to better capture user preferences. The models were evaluated using ranking metrics such as AUC, MRR, nDCG@5, and nDCG@10. Results show that contextual approaches significantly outperform non-contextual methods, with Disjoint LinUCB achieving the best performance. The project highlights the importance of feature engineering and contextual modeling in improving recommendation quality.