Accommodating Perturbation of Cluster Memberships in Optimal Trend by Multi-model Evolutionary Clustering

Evolutionary clustering based on differential evolution (deEC) is an algorithm to capture common similar data objects in each snapshot when the data objects change between the snapshots as a trend in the time-evolved dataset. deEC models this trend by smoothing any fluctuation on the changes of clusters between any two consequence snap-shots. It is argued that only yielding one model is insufficient to model the optimal trend since fluctuation of the change of clusters due to the optimal trend is also possible to occur. Smoothing this fluctuation accidentally deviates the model of the trend from the optimal one, in which this error caused in the previous snapshot cannot be corrected. In this paper, an improved work of the evolutionary clustering based on a multi-objective optimization approach is proposed as Multi-Model Multi-Objective Evolutionary Clustering (MM-MOEC). Multiple models of the trend are archived to capture different ways of smoothing the fluctuation. The models which cannot maintain this smoothness are discarded from the archive. The experimental results show that MM-MOEC outperforms deEC under several synthetic and real-life datasets.