Astudy On High-dimensional Ising Model Structure Learning And Itsapplications

The problem of structure learning for graphical models has attracted considerable attention. Traditionally, score-based approaches and search-based approaches are two main tools in learning a graph's structure. As a result of the promotion of science and technology, many fields are bombarded with high-dimensional data set. These traditional structure learning techniques become in-adequate due to a heavy combinatorial search when p>>N. Recently a neighborhood based approach draw the academic community's attention: this method is high-dimensional in nature and is model selection consistent under mild assumptions. Due to the performance of the penalty term, the objective function is convex but not differentiable, the optimization procedure of such objective function is computationally challenging. In this article, we proposed an efficient algorithm: coordinate gradient descent, first used by Friedman in solving generalized linear models with elastic net penalty, to tackle with the computational problem mentioned above. Then we applied the proposed method to analyze the underlying graph structure of a real data set.