IRBM: Incremental Restricted Boltzmann Machines for Concept Drift Detection and Adaption in Evolving Data Streams

In today’s dynamically evolving data landscapes, detecting and adapting to concept drifts in streaming data is imperative. Concept drift occurs when there’s a shift in the statistical characteristics of input features, like their mean or variance, or when the relationship between these features and the target label changes over time. This drift can decrease a model’s accuracy because the model is trained on older data. As the data evolves, the model becomes outdated, which can lead to incorrect predictions and reduced performance. This paper introduces the Incremental Restricted Boltzmann Machine (IRBM), an approach designed to address these challenges. The IRBM adapts the traditional architecture and learning paradigms of Restricted Boltzmann Machines (RBMs) to incrementally process and learn from evolving data streams, ensuring model efficacy and accuracy over time. Through extensive experiments, we demonstrate the IRBM’s ability to swiftly detect concept drifts, adapt its internal representations, and maintain robust performance even when confronted with significant data evolutions. The proposed approach outperforms existing methods with an accuracy of 77.42%, 75.32%, 92.12% and 89.21% for electricity, phishing, weather, and rotating hyperplane respectively. Our findings suggest that the IRBM not only offers an effective approach to understanding and adapting to changing patterns in streaming data but also outperforms the other state-of-the-art techniques.