Constructing networks and learning node features from noisy data

Abstract

Over the past several decades, there has been a rapid increase in the production of data from scientific experiments and human behaviors, thanks to advancements in technology. The data, which is often collected in various forms and may contain errors, can be effectively analyzed using networks that store both node interactions and properties. The accuracy of network structure estimation plays a crucial role in understanding complex systems and downstream research, such as drug discovery, online commodity recommendations, and gene function analysis. This dissertation contributes to the current understanding of network structure by investigating various network estimation approaches using heterogeneous node interaction and activity data. First, we propose a maximum a posteriori (MAP) estimation approach to reconstruct a target layer in a multilayer network. In this work, we have improved the SimHash algorithm to compute similarities between layers in multilayer networks. The layers that are most similar to the target layer are used to determine the parameters of the conjugate prior. This model allows us to predict missing links and direct experiments to identify nodes with similar functions. We evaluate the approach using two real-world multilayer networks, and the results demonstrate that the MAP estimation is effective in reconstructing the target layer, even when there are a large number of missing links. Second, to capture the correlation between nodes, we establish a method for constructing a gene co-expression network for the Anopheles gambiae transcriptome from 257 unique studies obtained with different methodologies and experimental designs. We introduce the sliding threshold approach to select node pairs with high Pearson correlation coefficients. The resulting network, which we name AgGCN1.0, is robust to random removal of conditions and has similar characteristics to small-world and scale-free networks. Analysis of network sub-graphs revealed that the core is largely comprised of genes that encode components of the mitochondrial respiratory chain and the ribosome, while different communities are enriched for genes involved in distinct biological processes. Building upon the network construction approach, we introduce GeCoNet-Tool, a comprehensive gene co-expression network construction and analysis tool. The network construction part of the tool offers multiple options for processing gene co-expression data obtained using different technologies. The output is an edge list with the option to assign weights to each link. In the network analysis part, users can generate a table that includes various network properties such as communities, cores, and centrality measures. Finally, to identify distance relationships between nodes, we propose an unsupervised learning framework that learns node vectors and constructs networks from noisy node activity data. The framework involves designing a scheme to generate random node sequences from node context sets, which are derived from raw node activity data. We utilize a three-layer neural network to train the node sequences and obtain node vectors, which enable us to construct networks and identify nodes with synergistic roles. Additionally, we propose an entropy-based approach to select the most relevant neighbors for each node in the resulting network. The effectiveness of our method is validated using both synthetic and real-world data. To summarize, this dissertation focuses on developing computational models to extract network structures from heterogeneous data. The models are suitable for the analysis of both node interaction and activity data. The results of this study are expected to enhance our understanding of network structures and benefit the study of complex systems with rich but noisy data.