Density Peak Clustering Algorithm and Optimization Based on Measurements of Unlikeness Properties in Position Sensor Environment
Zhe Yao, Kun Gao · IEEE Sensors Journal · 2021
With the increasing popularity of position sensors, the rapid development of the mobile Internet and increasingly high quality of communication facilities based on 5G networks, all walks of life are producing moving object trajectory data at an increasingly rapid rate. The density peak clustering algorithm (DPC) is an effective method based on the attributes of local density and relative distance.DPCcan find a peak density by identifying the clustering center using a decision graph without specifying the quantity of clusters beforehand and can identify clusters with arbitrary shapes. Unfortunately, because the computations of local density and relative distance only rely on the likeness matrix based on distance measurements,DPCresults are unsatisfactory in the following cases: 1) when the data dimension is high, and the distribution is uneven; 2) no uniform measurement exists for the computation of local density, requiring different degrees to be selected based on different data sets; and 3) the measurement of the truncated distance${d}_{c}$focuses on global data and disregards local information, allowing variations in${d}_{c}$to affect the result. This paper proposes an optimized density peak clustering algorithm based on measurements of unlikeness properties (UDPC) to solve these problems.UDPCuses the block-based unlikeness measurement method to compute the likeness matrix, determines k-nearest neighbor information based on the newly formed likeness matrix, and determines the local density measurement method based on the k-nearest neighbor information. Experiments on classic datasets show that the density peak clustering algorithm based on measurements of unlikeness properties performs better than theDPC,FKNN-DPCandDPC-KNNalgorithms. The proposed algorithm unifies the measurement of local density and mitigates the effects of the truncated distance${d}_{c}$on clustering.