Neighbourhood clustering for enhancing domain-adaptive zero-shot learning and beyond

Abstract Domain-adaptive zero-shot learning is an emerging and challenging task that extends zero-shot learning to scenarios where the source and target domains have different distributions. To address this issue, existing methods typically rely on category prototypes learned from the semantic embeddings of class labels and leverage pairwise semantic relationships to align the source and target samples in a shared feature space. While promising, these methods still face two key limitations: (1) they focus solely on pairwise relationships, limiting their ability to capture the complex, high-order structural correlations among category prototypes; and (2) they align samples only with the category prototypes, overlooking the intrinsic correlations among samples, which undermines alignment efficacy. To tackle these issues, we propose a new method, neighbourhood clustering for enhancing (NCE) learning framework, which fully exploits correlations among both category prototypes and samples. For category prototypes, we adopt a hypergraph-based approach to capture high-order correlations that go beyond simple pairwise relationships. During the alignment process, we incorporate both intraclass and interclass correlations among samples. Experimental results on the I2AwA and I2WebV datasets demonstrate that our method significantly outperforms state-of-the-art methods in terms of performance. Furthermore, to validate the effectiveness of our method in more challenging scenarios, we use it in underwater image scenarios. Experimental results show that our method significantly improves the accuracy and robustness of underwater image recognition.