Abstract: Existing online class imbalance learning methods fail to achieve optimal performance because their assumptions about enhancing minority classes are hard-coded in model parameters. To learn the model for the performance measure directly instead of using heuristics, we introduce a novel framework based on a dynamic EA called Online Evolutionary Cost Vector (OECV). By bringing the threshold moving method from the cost-sensitive learning paradigm and viewing the cost vector as a hyperparameter, our method transforms the online class imbalance issue into a bi-level optimization problem. The lower layer utilizes a base online classifier for rough prediction, and the upper layer refines the prediction using a threshold moving cost vector learned via a dynamic evolutionary algorithm (EA). OECV benefits from both the efficiency of online learning methods and the high performance of EA, as demonstrated in empirical studies against state-of-the-art methods on thirty datasets. Additionally, we show the effectiveness of the EA component in the ablation study by comparing OECV to its two variants, OECV-n and OECV-ea, respectively. This work reveals the superiority of incorporating EA into online imbalance classification tasks, while its potential extends beyond the scope of the class imbalance setting and warrants future research attention. We release our code for future research.
Submission Length: Regular submission (no more than 12 pages of main content)
Changes Since Last Submission: We added a subsection 3.4.5 to discuss the storage requirements of our method further.
Specifically, we provide a rationale for maintaining a certain range of storage and also suggest a potential solution.
Besides, we highlight how we fairly compared our method and baselines.
Code: https://github.com/t2ance/OECV
Assigned Action Editor: ~Lijun_Zhang1
Submission Number: 2884
Loading