Algorithm For Layer-specific Module Detection In Multi-layer Cancer Networks

The complex cancer system can be effectively modeled by network,and the operation mechanism and changes in systems can be fully explored by analyzing the network.Module detection is one of the research hotspots in network analysis and has attracted widespread attention.Among them,a module refers to a group of nodes that are closely connected in the network,while the nodes between different modules are sparsely connected.With the development of biotechnology and the increasingly interactive data,the multi-layer network constructed by integrating multiple types of data has become a research hotspot.Specific module detection is an extension of module detection in multi-layer network,which can be studied by clustering algorithm based on multi-layer network.Although an increasing number of specific module detection algorithm has been proposed,they are criticized for the low accuracy,sensitivity to the number of networks,incomplete network information.Therefore,to attack these issues and obtain comprehensive network structure information,two specific module clustering algorithm models are proposed.Layer-specific module detection algorithm based on non-negative matrix decomposition(LSNMF)and specific module detection algorithm based on joint adaptive graph learning and low-rank self-representation(jGLSM).The major contributions of this paper are summarized as follows:(1)The traditional algorithm uses the multi-layer network topology,which cannot fully characterize the multi-layer network and has low accuracy.LSNMF obtains more latent structural information of the network by exploring the features of vertices.LSNMF divides the feature matrix obtained by matrix decomposition into a common feature matrix and a specific feature matrix that obeys the specificity constraints to obtain specific information between multilayer networks.Specific information refers to the unique information that distinguishes the current layer from other layer networks.Moreover,to ensure that the specificity representation does not destroy the topology of the network,a locality preserving distance regularizer is introduced to balance the modularity and specificity of the structure.Finally,experimental results on artificial and real network datasets demonstrate the potential applicability of the proposed method.The proposed algorithm can explore more latent structural information and improve the clustering accuracy.(2)Although LSNMF improves the clustering accuracy,there are still some drawbacks:First,locality preserving distance regulari zation term utilizes the static adjacency relationship,which will contain some noise;Second,the structural information of the network cannot be obtained by matrix decomposition.To overcome these problems,jGLSM employs a joint learning strategy to integrate procedures of low-rank representation and graph adaptive learning.jGLSM obtains the subspace clustering structure of the network through low-rank constraints,imposes Hilbert-Schmidt independence(HSIC)constraints on the representation matrix,obeys specific information constraints on the subspace structure,and integrates the advantage of adaptive graph learning.Experimental results demonstrate that jGLSM significantly outperforms the state-of-the-art in terms of accuracy,implying that the superiority of the proposed method.Moreover,the complexity,parameter sensitivity and convergence analysis of the algorithm are provided.LSNMF and jGLSM methods are effective exploration of specific module detection,which provide new ways and opportunities for the research of specific module detection.