Center’s working activity is currently focused on two projects

Deep Learning for Textual User Modeling and Recommendations

In the modern Web, with the advance of social interactions between users and full-scale data mining of all information related to users, user profiling has become a very important problem. In this context, user profiling means converting the recorded user behaviour into a certain set of labels or probability distributions that capture the most important aspects of the user that can be further used for making new recommendations, providing targeted advertisement and audience analytics and so on.

It turns out that user profiling can be significantly augmented with natural language processing. Much of what goes on in social networks has the form of a text, and one can use texts generated by the user him/herself such as ‘wall posts’ or ‘statuses’ to mine his or her interests and a few other characteristics, such as demographic information. Recent developments in the field of deep learning for natural language processing have led to state of the art models that operate in a basically unsupervised fashion and do not require much linguistic insight.

User Preference Prediction in Visual Data

The project is focused on the problem of inferring user interests using his/her photos in a mobile phone. The purpose of this project is to create the state-of-the-art real-time offline method suitable for Samsung mobile devices, so an important part of the project is to develop deep learning models for computer vision that will fit into mobile devices. In contrast to the classical image classification problem, where the objective is to maximize classification performance for individual images, here the objective is to learn an overall user-level image category distribution. In order to solve this problem, we first need to examine the relationships between different categories and deep features extracted from the user photos in the training phase. These learned relationships can then be used to predict distribution over different interest categories by offline mobile analysis of images posted by a new user.