Predicting Friendship Among Subjects Based on Common Activities

Abstract

The digital traces of our online behavior tell lot about our acquaintances or friendship as well as how social circle is evolving or changing with time. I have examined the outcome of common activities like following common celebrity profile on twitter. In light of this I have analyzed a twitter follower and followed relationship data. In this we have data about celebrities followed by the user as well as data about the friendship of user. I propose a learning model based on conditional random fields, to predict the relationships between individuals based on common celebrity profiles they follow. I have tried to see the effect of all combined features (following different type of celebrities) and in combination with the online common activities based features. I have trained the model using subset of data and used it to predict the friendships on the other parts of the network. Different celebrities from different areas are considered like authors, footballers etc. Undirected Graphical models and conditional random fields are used as a base for this problem. Model is taken as a complete graph of relationships. UGM library of Matlab is used to train the model with various instances, each consists of randomly picked 150 relationships and feature set corresponding to each relationship. As a result we get parameterized feature vector which is used while testing another unknown instance. To see the accuracy I have plotted the ROC curve. Various other scenarios are listed under which we can make changes to the variables and see the effect.