Top-K Permutation-Based Feature Selection with Reused One-Hot Subspace and Threshold Optimization for University Student Telecom Churn Prediction
Year: 2026
Type: conference-paper
Abstract: As competition intensifies in an increasingly saturated mobile market, accurately predicting churn among university students with distinctive usage and switching behaviors is important for effective retention strategies. However, survey-based student datasets often contain high-cardinality categorical fields, and one-hot encoding expands them into a high-dimensional sparse space that may introduce redundant indicators and increase computational overhead. Building on our earlier work on importance-driven feature selection, this article refines the pipeline to provide a more compact, efficient, and deployment-friendly solution for sparse one-hot survey data. Within each cross-validation fold, the proposed workflow trains a Random Forest on the full one-hot training space, computes permutation importance using only the training data, and retains a controllable Top-K set of influential one-hot indicators. The selected feature subsets are evaluated through direct column-subspace selection from a fixed one-hot matrix, while an optional lightweight threshold search is performed on training-set predicted probabilities to support decision-oriented metrics. Experiments using stratified five-fold cross-validation on a customized student survey dataset show that the proposed Top-K permutation workflow maintains competitive predictive performance while substantially reducing feature-subspace evaluation time and providing explicit control over the number of selected indicators.
Authors: Jiejun Li, Norriza Hussin, Fengyu Fei, Gong Jing
Institutions: SEGi University, Jiangmen Polytechnic
Language: English
Subfield: Marketing
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W7220720184