Improving density peak clustering on multi-dimensional time series: rediscover and subdivide

The density peak clustering (DPC) algorithm identifies patterns in high-dimensional data and obtains robust outcomes across diverse data types with minimal hyperparameters. However, DPC may produce inaccurate pattern sizes in multi-dimensional datasets and exhibit poor performance in recognizing similar patterns. To solve these issues, we propose the rediscover and subdivide density peak clustering algorithm (RSDPC), which follows three key strategies. The first strategy, rediscover, iteratively uncovers prominent patterns within the existing data. The second strategy, subdivide, partitions patterns into several similar subclasses. The third strategy, re-sort, rectifies errors from the preceding steps by incorporating critical distance and nearest distance considerations. The experimental results show that RSDPC is feasible and effective in synthetic and practical datasets compared with state-of-the-art works.