| Clustering is one of the most common exploratory data analysis techniques for intuitive understanding of data structures.It can be defined as the task of identifying subgroups in the data,so that the data points in the same subgroups(clusters)are very similar,while the data points in different clusters are very different.Clustering analysis is based on the feature to find the subgroup of the sample,or to find the subgroup of the feature based on the sample.The introduction of differential privacy technology into clustering analysis is an unavoidable hot spot in the current research field.Differential privacy is a kind of data distortion technology,which can resist attacks under any background knowledge,and is not limited by the size of the data set.In the field of data mining such as clustering analysis,differential privacy technology can effectively reduce the exposure of personal privacy,so the research of differential privacy protection algorithm is of great significance.The difficulty is to obtain the best performance data mining model on the premise of ensuring data privacy.The research focus of this paper is how to better protect the privacy of clustering algorithms and improve the availability of corresponding clustering algorithms,including:To address the privacy leakage problem of predictive recommendation algorithm in the process of clustering analysis,a differential privacy-preserving quality-of-service predictive recommendation algorithm based on an exponential mechanism is proposed,which is based on an improved coverage algorithm to predict and recommend the default Qo S values of the target users based on similar users.In this paper,we both analyze the algorithm to satisfy ε-differential privacy through theory and demonstrate experimentally that the algorithm significantly improves the quality of service prediction accuracy.To address the lack of privacy in the whole process of k-means++ algorithm,a differential privacy DPk-means++ clustering algorithm based on Laplace mechanism is proposed,and to further improve the usability of the algorithm,an efficient DPk-means-ev algorithm,which can improve the selection of initial centroids and avoid the blindness in setting the k-value and the sensitivity in selecting the initialized centroids,and the experiment proves that the algorithm effectively improves the efficiency and usability of clustering.To address the problem of non-convergence in the practical use of the k-means algorithm,a new framework for injecting differential privacy into the real center of mass in an interactive environment is studied and proposed,where the center of mass motion is controlled directionally during the iteration of the clustering process,and then noise is injected to achieve convergence.The key properties of the method are verified theoretically,and it is shown to have at most twice the number of iterations as Lloyd’s k-means algorithm.Experimental results show that the algorithm has better clustering quality than existing differential privacy algorithms in an interactive setting while ensuring convergence with the same differential privacy guarantee. |
