We have advanced the state of the art in extracting user-level interests and community-level dynamics from social platforms. Our work spans users’ explicit, implicit and future interests by semantically analyzing their historical social content. This enables the prediction of users’ future interests and has applications in targeted recommendations, and content curation. We have also explored the challenges associated with identifying life events on social networks, which is extremely difficult, as life events are: (1) rarely shared and hence suffer high-class imbalance, and (2) mentioned with differing terminology. We have developed prediction-based models and encoder-decoder architectures based on Recurrent Neural Networks (RNN), making it possible to predict social behavior before it is reflected on a social platform.
On a macro level, We have established techniques based on multivariate time series analysis, neural embeddings, and causality analysis to identify highly like-minded groups of users that share similar temporal interests. This line of work shows significant improvements over the current state-of-the-art techniques in application domains such as news recommendation, user prediction and community selection. We have also explored the concept of link prediction in social networks, where the social network is viewed as a dynamic heterogeneous graph that is evolving over time, and shown improved performance over the current state-of-the-art methods. This work has wider applications, and can be used in other areas that model information in temporal graphical form. This work received Best Paper Award at ECIR 2019. Recently, We have shown how learning latent representations of dynamic heterogeneous graphs can enable the prediction of future communities on social networks.