A Deep Learning and Knowledge Transfer Based Architecture for Social Media User Characteristic Determination

Abstract

Determining explicit user characteristics based on interactions on Social Media is a crucial task in developing recommendation and social polling solutions. For this purpose, rule based and N-gram based techniques have been proposed to develop user profiles, but they are only fit for detecting user attributes that can be classified by a relatively simple logic or rely on the presence of a large amount of training data. In this paper, we propose a general purpose, end-to-end architecture for text analytics, and demonstrate its effectiveness for analytics based on tweets with a relatively small training set. By performing unsupervised feature learning and deep learning over labeled and unlabeled tweets, we are able to learn in a more generalizable way than N-gram techniques. Our proposed hidden layer sharing approach makes it possible to efficiently transfer knowledge between related NLP tasks. This approach is extensible, and can learn even more from metadata available about Social Media users. For the task of user age prediction over a relatively small corpus, we demonstrate 38.3% error reduction over single task baselines, a total of 44.7% error reduction with the incorporation of two related tasks, and achieve 90.1% accuracy when useful metadata is present.