MNERLP-MUL: Merged node and edge relevance based link prediction in multiplex networks

In multiplex networks, nodes can have multiple types of relationships (links) encoded into different layers such that each layer represents a single type of link. Even though the nature of links in different layers may differ, the nodes themselves remain the same, and so do their underlying relations among themselves. Combining the information in all the layers into one single network such that link prediction can be performed using all the available information is an ongoing research problem. In this work, we theorize that to accurately perform this link prediction, we have to take into account the relevance of both the edges as well as the nodes that connect two directly unconnected nodes. First, we utilize an aggregation model that encodes the information from different layers into one summarized weighted static network, taking into account the relative density of the layers themselves. Then, we propose an algorithm, MNERLP−MUL, which first calculates node and edge relevance based on the summarized graph, and then we combine both these factors to perform link prediction on unconnected pairs of nodes. The edge relevance is calculated using the information from the immediate vicinity of the edge (local information), while node relevance is calculated based on the node’s importance to the overall structure of the graph (global information). We use this methodology to model our method on quasi-local link prediction approaches, which attempt to inculcate properties of both local and global properties for increased accuracy. We compare our method with classical link prediction methods for weighted graphs, and the results indicate its superior performance, both on the summarized weighted graph and original layers.