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

Researchers introduce Neural Conjugate Aggregation Model (NCAM), a Bayesian framework for combining multiple biased sensor measurements without ground-truth labels. The method decomposes uncertainty sources and provides calibrated prediction intervals, with applications to sensor networks and scientific monitoring systems.

NCAM addresses a fundamental challenge in multi-source data fusion: aggregating measurements from sensors with unknown biases when labeled training data is unavailable. This scenario is common in environmental monitoring, IoT networks, and scientific research where ground truth is expensive or impossible to obtain. The framework solves this by modeling source-specific bias and reliability as context-dependent parameters, enabling the system to learn which sensors are trustworthy under different conditions.

The technical contribution combines neural networks with conjugate Gaussian inference, leveraging the analytical tractability of Bayesian conjugate priors. By decomposing total uncertainty into epistemic (reducible) and aleatoric (irreducible) components, NCAM provides interpretable uncertainty quantification rather than black-box confidence scores. The researchers resolve non-identifiability—a fundamental problem where multiple parameter combinations fit the data equally well—through sensor anchoring and variance regularization, ensuring stable and meaningful posterior estimates.

The integration of Monte Carlo conformal prediction with Bayesian inference is particularly valuable, as it provides finite-sample coverage guarantees without distributional assumptions. This bridges theoretical guarantees with practical uncertainty quantification. Experimental validation on air-quality datasets shows competitive performance against established baselines including Kalman filtering and probabilistic PCA.

For practitioners in industrial IoT, environmental monitoring, and scientific computing, NCAM offers a principled approach to heterogeneous sensor fusion without resorting to labeled data collection. The framework's ability to learn sensor-specific biases could reduce deployment costs and improve reliability in distributed sensor networks. Future work likely involves scaling to high-dimensional sensor arrays and handling non-Gaussian noise distributions.

• →NCAM enables unsupervised fusion of biased sensors by learning context-dependent source reliability without ground-truth labels
• →The framework decomposes uncertainty into epistemic and aleatoric components for interpretable confidence estimates
• →Structural non-identifiability is resolved through sensor anchoring, ensuring stable and meaningful posterior aggregation
• →Integration of conformal prediction provides finite-sample coverage guarantees alongside Bayesian uncertainty quantification
• →Experimental results demonstrate improved accuracy and calibration compared to mean aggregation, probabilistic PCA, and Kalman filtering