IEBench: Benchmarking Streaming Learners on Imbalanced Evolving Data Streams
Alessio Bernardo, Giacomo Ziffer, Emanuele Della Valle · 2021 International Conference on Data Mining Workshops (ICDMW) · 2021
Nowadays, data coming from electronic devices surround us. Analyzing them all in real-time, also considering the class imbalance and concept drift challenges is a big challenge and the starting point to understand where to innovate. This paper proposes IEBench, an easy-to-use benchmarking environment for comparing streaming learners. We implemented various state-of-art algorithms, and we conducted a comprehensive experimental campaign with IEBench. We evaluated the algorithms on artificial and real data streams with different imbalance levels and concept drift types. We collected empirical evidence of the role and impact of existing methods for rebalancing data streams in improving performance during different types of concept drift. IEBench eases the practitioners’ task of testing existing algorithms on a new data stream and the scientists’ one of developing new algorithms and systematically comparing them with the state-of-the-art.