Decentralized negative links prediction in signed social networks

Abstract

Negativity and conflicting thoughts are a part of users in social networks. To make our system grow in a positive direction, we should minimize the negative interactions among the users of our social networks. We have tried to predict the negative relationship among the users of a social network. The link prediction task is the process of detecting future links between two nodes of the graph whether there will be a positive link, negative link or no link between the selected edges. The negative link prediction task is the process of detecting future negative links between the nodes. We have used Epinions and SlashDot Zoo datasets to train our model and evaluate its performance. We have used a decentralized approach in which we divide our huge graph into multiple communities and handle the class imbalancing as we were dealing with a highly imbalanced dataset. There are very few negative links present in the dataset compared to the positive links due to which it is impossible to get accurate results by simply giving this graph as an input to the model. We need to preprocess the graph. We have extracted features from the graph using graph embedding techniques so that we can easily train our model on them and get accurate results. After getting features we trained our model and attached it to the backend of our web based application in which user enters two nodes and our application predict the link between those nodes with the help of trained model attached at the backend