Non-stationary data stream mining aims to classify large scale online instances that emerge continuously. The most apparent challenge compared with the offline learning manner is the issue of consecutive emergence of new categories, when tackling non-static categorical distribution. Non-stationary stream settings often appear in real-world applications, textit{e.g.}, online classification in E-commerce systems that involves the incoming productions, or the summary of news topics on social networks (Twitter). Ideally, a learning model should be able to learn novel concepts from labeled data (in new tasks) and reduce the abrupt degradation of model performance on the old concept (also named catastrophic forgetting problem). In this work, we focus on improving the performance of the stream mining approach under the textbf{constrained resources}, where both the memory resource of old data and labeled new instances are limited/scarce. We introduce the embedding model, which works on the class-embedding space from the encoder output, and continually constructs the prototypes to represent new categories without additional learned weight like the softmax classifier. We propose a simple yet efficient resource-constrained stream mining framework sysname{} based on prototype mechanism, it consists of two sub-steps: the contrastive-prototype learning and drift estimation. Specifically, the contrastive prototype learning is applied on unlabeled data to encode the semantically similar instances into an embedding space, then generate the discriminated prototype for each class. Next, during model updating on new tasks/categories, we implement a drift estimation strategy to compensate for the drift of each class’s prototype, which is used to reduce the knowledge forgetting without storing previous data. We perform experiments on public datasets (textit{e.g.}, CUB200, CIFAR100) under stream setting, our approach is consistently and clearly better than many state- of-the-art methods, along with both the memory and annotation restriction.

The Web Conference is announcing latest news and developments biweekly or on a monthly basis. We respect The General Data Protection Regulation 2016/679.