X-clustering beyond contextual representations

Tianyi Huang, Zhengjun Zhang, Xin Yuan, Stan Z. Li, Neal Naixue Xiong, Shenghui Cheng

Abstract: Despite the success of clustering and eXplainable AI (XAI), most clustering methods lack explainability. Traditional clustering generates pseudo labels but ignores probability distributions, cluster similarities, and local manifold structures, making the process and results difficult to interpret. Deep clustering improves this by jointly learning probability distributions and representations, but the learned representations often lack contextual meaning and misalign with data topology.To address this, we propose X-Clustering beyond Contextual Representations (XCR). It learns contextual representations and maps them for the visual explanation of the clustering probability distributions, cluster similarities, and topological structure of each cluster. Samples with higher cluster probabilities are positioned closer to corresponding anchors. Using an adjacency graph, connected samples stay near the same anchor with similar probabilities, while unconnected samples remain distant. Anchors are adaptively updated to index clustering distributions accurately. Finally, samples and anchors are fused into a unified space for intuitive visual explanations. Experiments demonstrate that compared with existing clustering methods, XCR not only provides better explainable visual clustering results but also achieves better average ACC and NMI on five testing datasets. Furthermore, the average sample-cluster tagging accuracy of anchors on five testing datasets is 97.3% with the visualization of UMAP.