The Regularization Methods With Group Structure And Link Prediction In Networks

With the rapid development of science and technology, high dimensional and massive data are produced in the national economy, biology, logistics and Internet industries. High dimensional data processing is one of the common problems faced by statistics, computer science and biological information science in recent years. Also the work of researches the structure characteristics is one of hot topics in all fields in the past decade.We study the variable selection problems with group structure information in the first part. We analyze several different sparse regularization methods and its framework. At the same time, we use the bound of the maximum of Gaussian random variables and other inequalities to research Group Lasso based on the model of Lasso, and prove the estimation consistency of the Group Lasso when the number of group variables is much larger than the number of samples.In the second part, we try to study the link prediction problem of social network by combining the structure information with similarity index and trans-forming un-weighted network to weighted one. The experiments show that struc-ture weight plays a vary important role in improving the prediction accuracy. Simultaneously, weighted RALP also performs better than both the weighted RA and weighted LP.