Adaptive Ensemble Thresholding for OOD Intent Detection

Out-of-domain intent detection in natural language understanding systems faces significant challenges from suboptimal threshold selection and signal degradation through inappropriate normalization techniques. This paper presents an adaptive ensemble thresholding framework that substantially extends our previous conference work by addressing fundamental limitations in existing variational autoencoder-based detection methods. Our approach combines reconstruction loss from variational autoencoders with classifier confidence scores to create a unified detection signal that captures both semantic deviation and prediction uncertainty. The framework incorporates a novel smart scaling strategy that preserves natural separation ratios between in-domain and out-of-domain samples, preventing the signal destruction caused by standard normalization approaches. Through systematic parameter optimization using grid search techniques, the method adaptively determines optimal ensemble weights and threshold selection strategies tailored to specific dataset characteristics. We evaluate our framework across multiple datasets with varying semantic complexity and domain structures, demonstrating consistent performance improvements over baseline variational autoencoder approaches and recent state-of-the-art methods. Compared to our previous VAE-based approach, the framework demonstrates an average performance gain of 3.15 percentage points across all evaluation metrics. Our analysis reveals that ensemble scaling strategy significantly impacts detection performance, with proper signal preservation being more critical than sophisticated threshold selection methods. This work provides a principled approach to adaptive ensemble learning for out-of-domain detection, offering a robust solution that generalizes effectively across diverse datasets and linguistic contexts including low-resource languages like Persian.