Language Predictor Analysis

This project focuses on predicting a voter’s primary language using voter record data containing 5.7 million rows, including around 160,000 labeled records and over 400 features, with the goal of improving multilingual voter outreach across Massachusetts. Last semester, students trained and deployed machine learning models on over 13,000 labeled records to classify languages such as Haitian Creole, Cantonese, Khmer, and Mandarin. Students last semester took two approaches: • Name-based modeling – Generated language likelihood scores for individual voters based on their names. • Precinct-based modeling – Language probabilities were inferred from precinct-level voting distributions, capturing geographic patterns of language use The outputs were combined into a dataset that appended probability scores for each target language, enabling outreach teams to identify which communities may benefit from multilingual materials. The goal of this project is to improve predictions for languages such as Cantonese and Khmer, as well as to combine the name- and precinct-based models into a single ensemble model. While the primary scope this semester is to build upon and refine the previous work, students are encouraged to explore and experiment with alternative approaches if they arise, and to discuss potential improvements with the client. It is recommended that students have some knowledge of regression models (using the scikit-learn package or similar), familiarity with feature engineering, and experience with the pandas package in Python.