.png){kind=icon}
.png){kind=icon}
Return
OTL-CE: Online transfer learning for data streams with class evolution
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2025.129470.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Learning from data streams has become a crucial area of research due to the continuous evolution of data sources. One of the key challenges is concept drift, where data distributions shift over time. However, real- world data streams not only experience concept drift but also undergo class evolution, where new classes appear, and existing ones vanish. Most existing approaches focus on adapting to concept drift, overlooking the possibility of class evolution in real-world scenarios, which limits their performance in practical applications. To address this issue, a class-based ensemble method, namely online transfer learning for data streams with class evolution (OTL-CE), is proposed. Upon the arrival of anew data chunk, OTL-CE employs a One-vs-All strategy to build a binary classifier for each class within the data chunk, subsequently adding it to the respective target sub-ensemble. To effectively transfer knowledge from heterogeneous source domains, a density-based correlation alignment method is introduced to align the distribution of data with the same class label. Following this, a source sub-ensemble is constructed based on the aligned source domain instances. Finally, knowledge from source and target domains is integrated by weighted combining the corresponding class sub-ensembles, where the weight of each sub-ensemble is determined by its prediction accuracy on the latest data chunk. The experimental results demonstrate that OTL-CE achieves superior performance across most datasets under four evaluation metrics: Accuracy, Balanced Accuracy, Recall, and G-mean. In particular, it outperforms other algorithms on the CovTYPE dataset, delivering a notable improvement of at least 5.8% in G-mean. Additionally, the results confirm that the improved density-based correlation alignment method can effectively alleviates negative knowledge transfer.
Keywords:
Transfer learning
Online learning
Data stream
Concept drift
Ensemble learning
Journal
IF:
6.5
Papers:
2.5W
Citations:
6.5W