OpenDAR: Distribution-aware reweighting for long-tailed open-world semi-supervised learning

Open-world semi-supervised learning (OWSSL) aims to discriminate known and novel classes. However, real-world data generally exhibits a long-tailed distribution, which causes the model to develop significant prediction biases toward the abundant head class samples. Existing studies always tackle the issues of OWSSL and long-tailed recognition in isolation, without considering the complex effects brought about by their coupling. When the imbalance of data distribution is intertwined with the uncertainty of novel classes, the absence of collaborative modeling for the overall class distribution leads to two critical challenges: representation collapse and confidence bias. To address these key limitations, this paper proposes an Open-world Distribution-aware Reweighting approach for long-tailed open-world scenarios, named OpenDAR, which collaboratively optimizes representation learning and pseudo-label refinement through a class distribution-aware reweighting mechanism. In OpenDAR, we propose a distribution-aware reweighted representation learning method, which employs a reweighting factor that focuses on tail and novel classes to guide the reweighted contrastive learning, thereby suppressing representation collapse. Furthermore, we design a debiased pseudo-label refinement method, which rebalances pseudo-labels according to the class distribution and selects high-quality ones with a long-tailed compensation threshold strategy, effectively mitigating confidence bias. Experiments on a series of benchmark datasets demonstrate that OpenDAR consistently outperforms state-of-the-art methods across various open-world long-tailed settings.