Loss Functions in Link Prediction Based on Graph Neural Networks

GNNs are neural networks that can be used to analyse graphs directly. We're going to discuss utilising GNNs for link prediction even though they can also be used for node-level tasks (like giving each website on the internet a score) and graph-level tasks (like predicting characteristics given a molecular structure). It is feasible to predict if a link will most likely connect two network nodes. Given the pervasiveness of networks, it has a wide range of uses, including reconstructing metabolic networks and recommending friends, movies, and knowledge graphs. To train these link prediction networks. We mostly use the binary categorical cross-entropy (CE) loss with adaptive moment estimation to train these link prediction networks. A loss function, cost function, or error function is a calculation formula that returns an event or the value of one or more variable to its true number, which in turn represents some “costs” associated with this event. In this study, we evaluate the effects of various loss functions on the effectiveness of GNNs for link prediction with specified labels and various sample sizes. Our findings for GNNs demonstrate that in both of the aforementioned circumstances, there is no discernible difference between models trained using CE and other conventional loss functions.