Connector 0.5: A unified framework for graph representation learning

Graphs are a universal language representing and visualizing relationships and connections between different entities or data points [1, 2, 3]. Graph structure widely exists in various practical application systems. For example, relationships between online users in social media could form a large social graph network. Another example is the recommendation system, where users' behaviours such as purchasing, browsing and rating products can be abstracted into an interaction graph between users and products. Graph representation learning methods aim to learn nodes and edges of graphs as low-dimensional vectors, mainly in Euclidean space [4]. These representations could then be used directly to improve various downstream tasks, such as node classification, link prediction, and visualization tasks. However, large amounts of data are still represented by different types of graphs, including homogeneous, heterogeneous, knowledge, and signed graphs [5]. Over the years, various graph embedding models have been proposed to transform graph entities into low-dimensional vectors [1, 2, 3, 6].