Concept drift adaptation with scarce labels: A novel approach based on diffusion and adversarial learning

The distribution of streaming data may change over time, making the knowledge previously learned by machine learning models outdated. This phenomenon is known as concept drift. It is common in many fields, such as weather forecasting, electricity prediction, and computer vision. To address this issue, concept drift adaptation is conducted by updating the parameters or adjusting the architecture of models. However, it may involve changing complex model structures and lead to slow convergence. This challenge is further exacerbated in scenarios with scarce labels, where instance labels are difficult to obtain and only a small amount of labeled data are available. The key to overcoming these challenges is to learn the good feature representations for incoming instances. Accordingly, we propose a novel concept drift adaptation method called CDAL, which incorporates diffusion learning into an adversarial network structure to polish up feature extraction under scarce labels. By simulating a graph-based diffusion process, diffusion learning enhances feature correlation among similar instances and weakens the feature correlation among different instances. On the other hand, adversarial learning enhances the feature extractor by guiding the latter to compete with a pair of specially designed classifiers. With these two techniques, CDAL can facilitate concept drift adaptation and ensure robust performance in scenarios with scarce labels. Additionally, our model exhibits strong performance despite its simple structure, enabling rapid adaptation in the presence of concept drift. We compared our method with four existing state-of-the-art methods on seven public datasets, including synthetic and real-world datasets. The results show that our method achieves superior performance across all datasets. •Diffusion learning can enable the model to learn better feature representations with a simple network structure.•A simple model structure allows for quick adaptation to concept drift, with the added benefit of requiring fewer labeled samples.•The adversarial learning can improve feature extraction capability and adapt to delayed label arrival.•Experiments with representative data sets verify the performance of the model.