Location-based services have attracted significant attention for the ubiquitous smartphones equipped with GPS systems. These services (e.g., Google map, Twitter) generate large amounts of spatio-textual data which contain both geographical location and textual description. Existing location-based services (LBS) assume that the attractiveness of a Point-of-Interest (POI) depends on its spatial proximity from people. However, in most cases, POIs within a certain distance are all acceptable to users and people may concern more about other aspects. In this paper, we study a region-aware top-k similarity search problem: given a set of spatio-textual objects, a spatial region and several input tokens, finds
most textual-relevant objects falling in this region. We summarize our main contributions as follows: (1) We propose a hybrid-landmark index which integrates the spatial and textual pruning seamlessly. (2) We explore a priority-based algorithm and extend it to support fuzzy-token distance. (3) We devise a cost model to evaluate the landmark quality and propose a deletion-based method to generate high quality landmarks (4) Extensive experiments show that our method outperforms state-of-the-art algorithms and achieves high performance.