H O P L P ? M U L: link prediction in multiplex networks based on higher order paths and layer fusion

Abstract

Multiple kinds of connections (links) may be encoded into distinct layers in multiplex networks, with each layer representing a particular type of link. Even if the type of linkages in various layers varies, the nodes themselves, as well as their underlying relationships, are retained. Considering the combined structure of all the layers, we achieve a complete overview of the network, which is impossible to achieve using any single layer itself. In this work, we theorize that this summarized graph (overview) provides us with an opportunity to determine the regional influence of nodes to greater certainty, and we can exploit this for more accurate link prediction. To begin, we use an aggregation model that combines information from many layers into a single summary weighted static network while accounting for the relative density of the layers. Then, we propose an algorithm HOPLP ? MUL which iteratively calculates link likelihoods taking longer paths between nodes into account. We also incorporate the concept of layer ranking based on densities as well as the dampening effect of longer paths on information flow. We compare our technique (HOPLP ? MUL) to stae-of-the-art multiplex link prediction algorithms, and the results show that it outperforms them both on the summarised weighted graph as well as the original layers. © 2022, The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature.