Overcoming Relational Learning Biases to Accurately Predict Preferences in Large Scale Networks

Many individuals on social networking sites provide traits about themselves, such as interests or demographics. Social networking sites can use this information to provide better content to match their users' interests, such as recommending scheduled events or various relevant products. These tasks require accurate probability estimates to determine the correct answer to return. Relational machine learning (RML) is an excellent framework for these problems as it jointly models the user labels given their attributes and the relational structure. Further, semi-supervised learning methods could enable RML methods to exploit the large amount of unlabeled data in networks. However, existing RML approaches have limitations that prevent their application in large scale domains. First, semi-supervised methods for RML do not fully utilize all the unlabeled instances in the network. Second, the collective inference procedures necessary to jointly infer the missing labels are generally viewed as too expensive to apply in large scale domains. In this work, we address each of these limitations. We analyze the effect of full semi-supervised RML and find that collective inference methods can introduce considerable bias into predictions. We correct this by implementing a maximum entropy constraint on the inference step, forcing the predictions to have the same distribution as the observed labels. Next, we outline a massively scalable variational inference algorithm for large scale relational network domains. We extend this inference algorithm to incorporate the maximum entropy constraint, proving that it only requires a constant amount of overhead while remaining massively parallel. We demonstrate our method's improvement over a variety of baselines on seven real world datasets, including large scale networks with over five million edges.