Neural Conjugate Aggregation: Identifiable Unsupervised Multi-Sensor Regression under Heterogeneous Sensor Bias

The article presents the Neural Conjugate Aggregation Model (NCAM), a hierarchical Bayesian framework designed for unsupervised multi-sensor regression in the presence of heterogeneous sensor biases and noise. NCAM utilizes neural networks combined with conjugate Gaussian inference to learn source-specific biases and reliability, producing analytically tractable posteriors that account for both epistemic and aleatoric uncertainty. The model's effectiveness is demonstrated through experiments on synthetic and real-world air-quality datasets, showing superior predictive accuracy and uncertainty calibration compared to traditional methods like mean aggregation and Kalman filtering, which is significant for practitioners dealing with sensor data fusion in uncertain environments.