Weighted Multi-view Clustering Method

Clustering analysis is fundamental technology in both data mining and artificial intelligence.Clustering can serve for the tremendous daily application like functional feature recommendation,user requirement analysis and so on.On account of the rapid development of mobile devices and sensors,there have been produced massive data.Data can be described by multiple views.The emergence of multi-view data,which promotes the traditional single-view clustering methods extend to the multi-view clustering methods.Although many multi-view clustering algorithms have been developed so far.However,most multi-view clustering algorithms only look at multiple views on average.Such an approach causes the algorithm to get less than optimal clustering because some noisecontaining views have the same weight as other views.Therefore,a new entropy-based viewweighting strategy is proposed in this paper.In order to recap the effectiveness of this strategy,combine it with robust multi-view k-means which based on l2,1 norm to get a new weighted multi-view k-means algorithm.The algorithm can dynamically assign weight to the views.Self-paced learning strategy which simulates human learning process is gradually introduced into machine learning area.In this paper,a new self-paced learning penalty item is proposed,which is combined with the traditional multi-view k-means and spectral clustering methods respectively.New self-paced learning based multi-view k-means algorithms and selfpaced learning based multi-view spectral clustering are proposed.Among them,the second algorithm is a new algorithm that combines self-paced learning strategy and spectral clustering method for the first time.The proposed algorithm can define the complexity of views according to view loss in the clustering process.And first learn from easy views and then gradually adding the complex views to the clustering task.The proposed weighted multi-view clustering algorithms are compared with several multiview clustering algorithms in multiple real world dataset.Experimental results demonstrate that the delivered algorithms are more efficient.