Improved User Identification through Calibrated Monte-Carlo Dropout

Rouhollah Ahmadian, Mehdi Ghatee, Johan Wahlström

Abstract: Highlights•Utilizes CNNs to capture local temporal dependencies by segmenting input data.•Applies Monte-Carlo Dropout during testing to quantify prediction uncertainty.•Implements sophisticated calibration techniques to address miscalibration issues.•Introduces Ensemble of Near Isotonic Regression as the efficient calibration method.•Leverages probabilistic decision fusion for enhanced accuracy.•Integrates results through uncertainty-based weighted averaging.•Rigorously evaluates proposed methods across multiple datasets.•Demonstrates significant improvements in user identification performance.•Optimizes Monte-Carlo sampling to enhance feasibility for real-time applications.