Margin-aware active learning for user-adaptive text classification

We propose a human-in-the-loop text classification framework for rapid personalisation of deep neural networks (DNNs) to user-specific preferences under strict label budgets. The core is a margin-preserving embedding aligned with a margin-aware active learning (AL) strategy that jointly optimizes uncertainty, diversity, density, and class coverage in query selection. Use of a fixed local embedding and lightweight linear SVM classifier ensures rapid model updates. To resolve the cold start problem we introduce a one-class initialisation with ring-radii bisection from a single user-selected example to bootstrap learning, along with a hybrid AL acquisition function that evolves from exploitative to exploratory querying. In experiments on a proprietary dataset (UKAC) and 20 Newsgroups, the proposed AL strategy consistently outperforms random sampling, achieving higher accuracy with far fewer labels especially in the early stages of labeling. Our approach yields superior label efficiency (Macro-F1) at small label budgets, demonstrating a practical solution for user-adaptive DNN personalisation. The code is available at https://github.com/tsantosh7/Margin-Preserving-Active-Learning.