A novel open-set clustering algorithm

DOS (Delta Open Set) is an interesting clustering algorithm that transforms cluster identification into set identification. It identifies the objects whose neighborhoods coincide as an open-set, and an open-set corresponds to a cluster. However, once the dataset is complex, DOS tends to identify overlapping clusters as one category. We believe the main reason is that DOS unifies the neighborhood radius by a specific function, resulting in the inability to cope with various object distributions. To improve DOS, we propose DOS-IN (Irregular Neighborhoods). Specifically, DOS-IN generates irregular neighborhoods based on the similarity between objects to self-adapt to diverse object distributions. As a result, DOS-IN not only can accurately distinguish overlapping clusters but also has fewer input parameters. In addition, DOS-IN introduces the small-cluster merging mechanism to address the shortcoming of DOS in recognizing Gaussian clusters. The experimental results show that DOS-IN is completely superior to DOS. Compared with baseline methods, DOS-IN outperforms them on 7 out of 10 datasets, with at least 13.8% (NMI) and 2.4% (RI) improvement in accuracy. The code of DOS-IN is available at https://github.com/Youth-49/2023-DOS-IN.