Accelerated Real-Time Classification of Evolving Data Streams using Adaptive Random Forests

Machine learning is increasingly applied to a wide range of real-time applications, with classification tasks playing a critical role in enabling intelligent decision-making. However, the phenomenon of concept drift, in which the underlying data distribution changes over time, presents a significant challenge for maintaining the accuracy of machine learning models in applications with evolving data streams, such as health monitoring or sensor data analysis. The Adaptive Random Forest (ARF) algorithm addresses this issue by coupling multiple Hoeffding Trees with a drift detector to adapt to concept drift. As training a forest of growing decision trees is a high-latency operation, custom-hardware acceleration is needed to meet the stringent latency requirements for real-time use of ARF. To the best of our knowledge, this work describes the first FPGA implementation of the ARF algorithm, focusing on achieving high hardware efficiency, scalability, and adaptability to different datasets. We present a parameterized design that incorporates various levels of parallelism, resource sharing, and pipelining, and delivers 15 x-79 x faster execution than a 40 -core CPU with a maximum accuracy loss of 13%. Furthermore, our design outperforms a state-of-the-art GPU implementation, achieving 3x-21 x faster execution while maintaining accuracy scores in the range of 0.3% to 15% of the GPU ARF implementation.