Online Confirmation-Augmented Probabilistic Topic Modeling in Cyber-Physical Social Infrastructure Systems

Online probabilistic topic models serve as essential analytical tools within Cyber-Physical Social Infrastructure Systems (CPSIS), enabling the analysis of real-time data streams. These models empower operators and decision-makers with actionable insights, anomaly detection, predictions, optimized resource allocation, user engagement, and social feedback, all critical for responding to evolving CPSIS conditions. While these models use inferred topic-assignment distributions to create lower-dimensional representations, applying them to online user-generated streams, like social media and community apps, has historically posed challenges due to sparse relevant content, leading to suboptimal performance. Our study proposes a novel and expanded version of topic models that integrates the variational lower bound with a linear reward function, supervised by a label associated with the confidence of relevant content presence. We introduce a learning algorithm designed for these augmented topic models. Our empirical experiments, conducted on real-world datasets, provide compelling evidence that our approach uniquely enhances the potential of any topic model in CPSIS for downstream tasks in information management. These enhancements encompass improved topic interpretability, enhanced data labeling precision, and the refinement of similarity metrics, reinforcing the effectiveness of our online confirmation-augmented probabilistic topic modeling approach in processing and analyzing CPSIS real-time data streams.