An Adaptive Multi-granular Clustering Model Based On Density Peaks

Complex tasks in the real world tend to have inherent hierarchies.It is a natural and intrinsic logic of artificial intelligence to simulate the hierarchical knowledge processing by human brains.The key is to study the relationship between the representation and generative strategy of information particles,the topological structure of particles and the correlation between different layers.The breakthrough point of my research is based on the Density Peak Clustering(DPC)algorithm because of its three important characteristics.Firstly,the method is novel,concise and clear,which is in line with the intuitive cognition of the human brain to summarize the knowledge,i.e.,computing without complex logics.Secondly,the the decision gragh of DPC inherently contains the centers of granules in various levels.It is easy to cluster the analogous samples through the hierarchical clustering method and form the information particle generation mechanism.Thirdly,the numerical knowledge has the uncertain distribution form in the dimensional space.However,DPC belongs to the class of density clustering and the result of the induction would not be affected by the distribution form.Therefore,it is significant to study on the multi-granularity knowledge representation and evolution model based on DPC.It would establish a two-way cognitive mechanism heuristic for complex tasks of big data.Based on DPC,this thesis studies a multi-granularity knowledge discovery model,and makes several researchs as the following:1.A granular tree(GT)structure based on DPC is proposed.It is focused on the multi-granularity decomposition mechanism,i.e.,from coarse to fine and top to down mechanism.Combining human prior knowledge,we divide the dataset into independent subsets of different informative particle sizes and establish the inter-link between the layers of grains.Through permutation and combination,you can reach the specific hierarchical result in the problem solving space according to your needs.The algorithm finally projects DPC from the original plane clustering to multi-granularity space,forming a cognitive model of multi-granularity.2.We research on the robustness of DPC,which mainly includes the study on the limitations of the decision graph and the situations of decision graph failure are summerized.Then we put forward a granularity merging strategy.On the adaptive particle fusional parameter,an adaptive merging threshold driven by data is proposed based on two resonalble hypothetical bases.The method could automatically update the value of the threshold iteratively in the granular transformation process which is competitive to the artificial specified value traditionally.3.Based on the model proposed in the second research point,this section explores the "triple" generation mechanism of the multi-granularity model using the optimization of step size iteration,and suggests the clustering conclusional layer by the frequency of the results.