Study On Structural Evolutionary Clustering Method Without Prior Parameters

Human beings conceptualize and abstract the information obtained from the external world through clustering,to achieve feature extraction and dimension reduction.Because clustering can form clustering knowledge,it has been widely used in data mining and machine learning.However,although the current clustering algorithm can complete the clustering task well,it still needs the help of human experience to form the clustering knowledge in related fields.Without the support of human experience,the current clus-tering algorithm is difficult to complete the relevant clustering tasks,or even can not work usually.As the primary method of unsupervised learning,clustering should pay more at-tention to how to start from 0 and form clustering knowledge like human beings,rather than simply completing the clustering task with the help of existing human knowledge.Therefore,this thesis mainly focuses on the clustering scene in the unknown and open environment,studies how to simulate human clustering mechanism through algorithm design,realize the construction of clustering model from 0,and gradually develop and improve the model through later learning,to make it form clustering knowledge indepen-dently.That is to design a structurally evolvable clustering algorithm,which can cluster new samples using the existing algorithm model without prior knowledge and learn from new samples to evolve the algorithm model.This thesis mainly uses two ideas to design an evolutionary clustering algorithm,and one is based on the density method,the other is based on the dynamic core method based on the memory saturation degree.Finally,two structurally evolvable clustering algorithms are verified in practical application.The main work and innovations mainly include the following three aspects:Firstly,a density-based evolutionary clustering algorithm(DBEC)is proposed.Be-cause of the problems that the current clustering algorithm needs to set the number of clusters manually,there is a phenomenon of forgetting,and the algorithm model cannot evolve.DBEC algorithm regulates the evolution of the model by designing a control pa-rameter that can change adaptively with the increase of data.By adjusting control param-eters by adding new samples,the algorithm can maintain the stable memory of previous learning results without a priori parameters and modify the clustering results by adding new samples.Furthermore,based on the combination of different evolutionary strategies,we study evolutionary DBEC algorithms with different clustering characteristics,mainly including conservative DBEC algorithm,robust DBEC algorithm,and radical DBEC al-gorithm.Finally,simulation experiments verify the effectiveness of different types of DBEC algorithms.Secondly,based on the robust DBEC algorithm,a clustering algorithm based on evolution and propagation patterns(EPC)is proposed,which can effectively improve the efficiency of the algorithm without losing the clustering accuracy.In the evolution stage,the EPC algorithm extracts a small amount of data through random sampling in the sample space for evolution to obtain the initial cluster number and clustering results.The unla-beled data can be clustered into labeled data according to the nearest neighbor principle in the propagation stage.Through comparative experiments and ablation experiments,it is verified that the EPC algorithm has excellent clustering performance and robustness.Thirdly,a dynamic core evolutionary clustering algorithm(DCC)based on memory saturation is proposed.Because the number of clustering cores of K-means and other al-gorithms is often set before implementing the algorithm,it is difficult to search the appro-priate sample cluster through the increase or decrease of the number of clustering cores.Therefore,this thesis proposes an evolutionary algorithm DCC,in which the number of clustering cores dynamically adjusts with the change of data.DCC algorithm uses Gaus-sian function as the activation function of each dynamic core,and each core searches the sample clusters in the space by adjusting the center vector and coverage.The algorithm evolves from an initial core,and splits the winning core or adjusts its parameters through competitive learning after each new sample is added,to realize the evolution of the al-gorithm model.Finally,experiments on datasets show that the evolutionary clustering algorithm based on the dynamic core method has excellent clustering performance and strong robustness.Finally,two methods of evolutionary clustering are experimentally studied in practi-cal application scenarios.When clustering the target image in a video stream,most of the current clustering algorithms often fail when there is a lack of prior knowledge and too few samples.Therefore,based on the analysis of the urgent demand for an evolutionary clustering algorithm in a video scene,we apply the DBEC to the face clustering task in a video,and apply the DCC to the pipeline task to cluster geometric objects.