PROJECT TITLE :

Line Graph Neural Networks for Link Prediction

ABSTRACT:

We take a look at the graph link prediction task, a classic example of a graph analytical problem that has many practical applications in the real world. Current methods for link prediction typically compute features from subgraphs that are centered at two neighboring nodes and then use the features to predict the label of the link that exists between these two nodes. These advancements in Deep Learning have made this practice possible. A problem involving link prediction is transformed into a challenge involving graph classification by this formalism. In the deep learning model, graph pooling layers are required in order to extract fixed-size features for classification. As a result, there is information loss as a result of this requirement. We propose making use of the line graphs that are found in graph theory in order to overcome this significant limitation. This will require us to search for a path that is completely unique and unexplored. To be more specific, each node in a line graph corresponds to a different edge in the graph that was originally created. As a result, link prediction problems in the original graph can be solved in an equivalent manner as a node classification problem in the line graph that corresponds to it, as opposed to a task that involves graph classification. The results of our experiments on fourteen different datasets derived from a variety of applications demonstrate that our proposed method routinely outperforms the methods that are considered to be state-of-the-art, despite having fewer parameters and a high level of training efficiency.