Research On Evolutionary Clustering

Time-evolving data is ubiquitous in many dynamic scenarios, which refers to a collection of data that evolves over the time. Learning on time-evolving data has become a new important topic in the field of machine learning and data mining. And that clustering on time-evolving data is of the special significance which, compared with the supervised learning, is attracting increasing attention.The task of clustering on time-evolving data is that, when new data arrives, the system produces a clustering for each time step. The essential problems of evolutionary clustering include how to ensure the clustering at each time step remain faithful to the current data as much as possible, simultaneously not deviate too much from the recent past, and how to perform the temporal smoothness. These problems are also the major studies in this paper. Specifically, the research work in this paper consists of the following aspects:(1) An evolutionary clustering method (deEC) based on DE is proposed. By taking the advantages of evolutionary algorithms, the proposed method in this paper modifies the parameter a in temporal smoothness framework, so that it can adapt itself at each iteration of DE. And this helps the algorithm more quickly reach the balance between the clustering quality and temporal smoothness. In deEC, the temporal smoothness is expressed as the adaptability of an individual in the old environment. The higher fitness value of an individual in the old environment indicates the clustering yielded by the individual fit the historical data better. And the clustering are considered more consistent in the successive time steps. The experimental evaluation are conducted, using the synthetic data set and real world data set to clarify the proposed method.(2) The evolutionary clustering problem is researched from view of multimodal optimization. Most of the existing evolutionary clustering methods are based on the temporal smoothness framework, which ensure the temporal smoothness by adding a penalty term to penalize those clustering that deviate too much from the recent past. From view of multimodal optimization, we adopt a multimodal optimization method to perform a global/local search, and use a selection strategy based on NMI to select the best solution for the current time step. Using the synthetic data set and real world data set, the experimental evaluation are conducted, and the characteristics of the proposed method is analyzed.(3) The temporal smoothing is performed at the data level. Most of the existing evolutionary clustering methods perform the temporal smoothing at the model level. In this paper, two different techniques are proposed which perform the temporal smoothing at the data level. These two techniques make use of the information about the relationships between the historical data to build the current data matrix, and adopt the hierarchical clustering methods to obtain the final clustering result. The experimental evaluation are conducted using the synthetic and real world data set. The proposed methods are compared with the existing methods and their performances are analyzed.Evolutionary clustering, as a new research topic, has attracted increasing attentions in recent years. How to ensure the clustering at each time step remain faithful to the current data as much as possible, simultaneously not deviate too much from the recent past, and how to perform the temporal smoothness, are the essential problems. Therefore, in this paper, an evolutionary clustering based on DE is proposed, then from view of multimodal optimization, another evolutionary clustering is proposed, and the temporal smoothing is performed at the data level. We conducted the experimental evaluation and the analyzed the performance of the proposed methods. The work presented on this thesis has a guiding significance for the research on the evolutionary clustering.