Cost-Effective Clustering by Aggregating Local Density Peaks

Wen-Bo Xie, Bin Chen, Jun-Hao Shi, Yan-Li Lee, Xin Wang, Xun Fu

Abstract: Hierarchical clustering algorithms that provide tree-shaped results can be regarded as data summarization and thus play an important role in the application of knowledge discovery and data mining. However, such structured result also brings a challenge, i.e., a difficult trade-off between complexity (time and space) and quality. To tackle of this issue, we propose a newly designed agglomerative algorithm for hierarchical clustering in this paper, which merges data points into tree-shaped sub-clusters via the operations of nearest-neighbor chain searching and determines the proxy of each sub-cluster by the process of local density peak detection. Extensive experimental studies on real-world and synthetic datasets show that our method performs well by outperforming other baselines in accuracy, response time, and memory footprint. Meanwhile, our method can scale to half a million data points on a personal computer, further verifying its cost-effectiveness.