Methods of network embeddings and their applicationsMakarov, Ilya (Avtor) Šubelj, Lovro (Mentor) graph embeddingsinformation fusionnode classificationlink predictionnode clusteringcommunity detectiongraph visualizationnetwork scienceMany systems in the real world can be modeled as graphs or networks. One of the main goals of graph modeling is to formulate a general technique able to process structural data including relations between objects, which may also have some domain-specific information. In this thesis we deal with network representation learning, aiming to automatically learn low-dimensional vector features for the simplest graph motifs, such as nodes and edges. We search for a way that would enable efficiently solve machine learning problems on graphs, which include node classification, link prediction, node clustering, and graph visualization. We start with experimenting on embedding operators for pairs of nodes over existing node embedding models to efficiently solve the link prediction problem. We formulate several constraints on the embedding of pair of nodes based on their local proximity and validate our findings in classification and regression settings via various experiments on link prediction problem. We interpret the obtained results on the binary classification of missing links in co-authorship networks as a recommender system for scientific matchmaking on a real-world HSE University dataset. Taking the idea of joint node-edge embeddings, we develop new techniques of embedding nodes and edges under joint constraints via autoencoding graph and related to it Line graph. We incorporate two graph autoencoders for the graph and its Line graph while adding structural feature engineering, Laplacian regularization, and training the model with the help of $\beta$-masking on graph reconstruction task. The core feature of the model is in joint constraints on node and edge embeddings in the graph and its Line graph that help our model outperforming existing baselines for the node classification and the link prediction problems in unsupervised and semi-supervised settings. Following the growth of research in the field of network representation learning, we provide an extensive overview of existing network embedding techniques based on different taxonomies, such as model construction ideas, network data types, application settings, and stated open problems in network representation learning. Our survey is accompanied by state-of-the-art models' performance comparison on benchmark datasets and random graphs, thus shedding light on which model should be chosen for which network data under node classification, link prediction, node clustering, and visualization frameworks. Focusing on one of the open research questions on attributed network representations, we study the problem of constructing combined embedding models for networks attributed with text data and provide an in-depth analysis of fusion models for text and graph information on benchmark datasets considering node classification, link prediction, and graph visualization problems. We show that despite not having one fusion model to outperform all the others, there is an opportunity to choose a simple model based on text and network properties to be on par with state-of-the-art graph neural network models. Overall, the thesis covers the rapidly developing field of network representation learning and provides a comparison of modern models performance in benchmark settings. We provide novel models of network representation learning and partially answer the question, which models under which training setting and network dataset may perform better. We also show applications of network embedding models in machine learning, recommender systems, and information fusion.20212022-01-03 11:53:53Doktorsko delo/naloga134243sl