Date of Award
2026
Document Type
Thesis
Degree Name
Master of Science (MS)
Department
Computer Science
Committee Chair
Tathagata Mukherjee
Committee Member
Vaidyanath Areyur Shanthakumar
Committee Member
Jacob Hauenstein
Research Advisor
Tathagata Mukherjee
Subject(s)
Computational linguistics, Machine learning, Classification--Data processing
Abstract
The performance of encoder-only text classifiers degrades silently as their input distribution drifts over time. This thesis presents a closed-loop pipeline for drift-aware continual fine-tuning that combines a sample size invariant drift severity metric, an automated retraining loop on the most drifted samples, and a validation gate that rejects unsafe updates before deployment. We evaluate the pipeline on two datasets, Amazon Electronics Reviews and HuffPost News Category, and observe significant performance improvements over the initially trained model, along with a substantial reduction in compute and labelling costs compared to fresh-per-year and cumulative retraining baselines.
Recommended Citation
KC, Simran, "Drift aware continual finetuning of encoder-only text classifiers under temporal distribution shift" (2026). Theses. 857.
https://louis.uah.edu/uah-theses/857