Mining the Temporal Statistics of Query Terms for Searching Social Media Posts

There is an emerging consensus that time is an important indicator of relevance for searching streams of social media posts. In a process similar to pseudo-relevance feedback, the distribution of document timestamps from the results of an initial query can be leveraged to infer the distribution of relevant documents, for example, using kernel density estimation. In this paper, we explore an alternative approach to mining relevance signals directly from the temporal statistics of query terms in the collection, without the need to perform an initial retrieval. We propose two approaches: a linear ranking model that combines features derived from temporal collection statistics of query terms and a regression-based method that attempts to directly predict the distribution of relevant documents from query term statistics. Experiments on standard tweet test collections show that our proposed methods significantly outperform competitive baselines. Furthermore, studies of different feature combinations show the extent to which different types of temporal signals impact retrieval effectiveness.