Research On Multi-task Clustering

Traditional clustering methods can only learn the data in a single task.However,in many cases,the data in a single task are too limited to achieve better clustering performance.In the real world,there are many related tasks,multi-task clustering improves the clustering performance of all the tasks by transferring related knowledge among tasks.The research about supervised multi-task learning methods that use labels is relatively mature,but there are many problems to be solved for multi-task clustering methods that do not use labels.There are three data types that multi-task clustering can deal with:same-domain data,multi-domain data and multi-view data.This dissertation designs multi-task clustering methods for these three types of data,respectively.(1)Multi-task clustering for same-domain data:same-domain data means that the multi-task data come from the same domain.Due to the drawbacks of the existing multi-task Bregman clustering method,this dissertation proposes three improved methods.? Smart multi-task Bregman clustering:this method introduces the loss function as the judgement condition to solve the centroid skewing issue of multi-task Bregman clustering.? Multi-task kernel clustering:this method maps the data into the reproducing kernel Hilbert space to deal with the issue that multi-task Bregman clustering cannot cluster nonlinear separable data well.?Smart multi-task kernel clustering:this method introduces the loss function and maps the data into the reproducing kernel Hilbert space to solve both the centroid skewing issue and the clustering issue for nonlinear separable data.(2)Multi-task clustering for multi-domain data:multi-domain data means that the multi-task data come from different domains.First,due to the issues that only a few partially related multi-task clustering methods are proposed and they have limitations,this dissertation proposes two generalized methods:? self-adapted multi-task clustering uses the shared nearest neighbor similarity to transfer the instance knowledge only between the subtasks that are constructed by the related clusters of the tasks;? multi-task model relation learning clustering transfers the model parameter knowledge by learning the relatedness of the linear regression model parameters of the clusters between the tasks.Second,due to the issue that existing multi-task clustering methods can only transfer one kind of feature,instance and model parameter knowledge so that the related knowledge among the tasks cannot be fully used,this dissertation proposes three methods that transfer both feature and instance knowledge:?feature and instance transfer based multi-task clustering is suitable to completely related tasks;? manifold regularized coding multi-task clustering is suitable to partially related tasks;?feature and instance transfer based weighted multi-task clustering is suitable to both completely and partially related tasks.(3)Multi-task clustering for multi-view data:multi-view data means that the data in each task contain the features from different views.So far no methods have been proposed for tackling this case.This dissertation first proposes a co-clustering based multi-task multi-view clustering framework,which consists of three parts:single-task single-view clustering,multi-view relationship learning and multi-task relationship learning.Then based on this framework,this dissertation proposes two methods:? bipartite graph based multi-task multi-view clustering uses bipartite graph co-clustering,which can only deal with the data with nonnegative features;?semi-nonnegative matrix tri-factorization based multi-task multi-view clustering uses semi-nonnegative matrix tri-factorization co-clustering,which can also deal with the data with negative features.