Unsupervised Out-of-Distribution Sample Selection for Reliable ML in Wireless Systems
Year: 2026
Type: conference-paper
Abstract: Incorporating Machine Learning (ML) approaches within wireless systems hinges on their ability to handle real-world variability. In practice, data will inevitably deviate from the training distribution due to new devices, changing environments, and unforeseen interference. Out-of-Distribution (OOD) detection provides a safeguard by identifying these shifts before they lead to unreliable model behavior. To address this challenge, we develop a multi-stage OOD detection algorithm and evaluate it on the publicly available 3GPP NR CSI Compression dataset, which contains high-dimensional Channel State Information (CSI) representative of realistic wireless propagation. First, we perform coarse OOD detection by measuring the distance of incoming samples to in-distribution class centers, flagging those that lie beyond a threshold. We then apply fine OOD detection using density-based clustering to filter spurious anomalies and isolate dense clusters that representing meaningful shifts. While these stages identify OOD samples, retraining on all of them is inefficient. Instead, we rank the selected OOD samples by discriminativeness (distance from the ID distribution) and diversity (feature dissimilarity), ensuring that retraining incorporates only the most informative examples. In our analysis, coarse detection led to a 6.93% increase in reconstruction Mean Squared Error (MSE), fine detection raised it to 37.6%, and incorporating discriminativeness–diversity ranking further increased it to 45.5%. While MSE is typically minimized during training, a higher reconstruction error for detected samples indicates stronger deviation from the learned manifold. This increase confirms that the detected OOD samples reflect realistic wireless variations, reinforcing the framework’s value for reliable ML deployment in dynamic environments.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7168280342