What Do Your Look-alikes Say about You? Exploiting Strong and Weak Similarities for Author Profiling

We describe a two-step procedure for author profiling, which first exploits language similarities between users and then aims at discovering more complex dependencies for dissimilar users. The method is motivated by the fact that authors using very similar vocabulary are likely to have similar traits. We use both word-based and text-based features, as well as relying on previous research. The proposed approach gives successful results, especially for gender and age prediction. Moreover, we show the most useful features using relevance measures based on random forests.