Improving Label Prediction in Social Networks by Adding Noise

Social Networks like Facebook and Linkedin have grown tremendously over the last few years. This growth translates to more users, more information about users and at the same time an increase in the amount of information missing about the users. Techniques like Label Prediction/Collective Classification, Link Prediction alleviate this lack of information by estimating or predicting the missing information and they make use of the structure of social network or relational data amongst other things. Artificially corrupting the training data by adding noise has been shown to improve prediction performance in text and images as noising acts as a type of regularization. In the past, this technique has been used primarily in deep learning systems as a way of preventing model overfitting. Although, recent advances in machine learning show its broader applicability to other models, this technique has still not been applied for noising in relational networks to improve prediction performance. In this thesis, we have proposed a new generic framework of adding noise to relational data that can be easily incorporated into the existing relational machine learning frameworks. We have shown with experiments on real data that adding noise improves the collective classification accuracy by reducing either the bias or the variance or both. We have also compared this technique with the state of the art collective ensemble classification techniques and showed that our method outperforms it significantly