SG-WSTD: A framework for scalable geographic web search topic discovery

Abstract

Search engine query logs are recognized as an important information source that contains millions of users’ web search needs. Discovering Geographic Web Search Topics (G-WSTs) from a query log can support a variety of downstream web applications such as finding commonality between locations and profiling search engine users. However, the task of discovering G-WSTs is nontrivial, not only because of the diversity of the information in web search but also due to the sheer size of query log. In this paper, we propose a new framework, Scalable Geographic Web Search Topic Discovery (SG-WSTD), which contains highly scalable functionalities such as search session derivation, geographic information extraction and geographic web search topic discovery to discover G-WSTs from query log. Within SG-WSTD, two probabilistic topic models are proposed to discover G-WSTs from two complementary perspectives. The first one is the Discrete Search Topic Model (DSTM), which discovers G-WSTs that capture the commonalities between discrete locations. The second is the Regional Search Topic Model (RSTM), which focuses on a specific geographic region on the map and discovers G-WSTs that demonstrate geographic locality. Since query log is typically voluminous, we implement the functionalities in SG-WSTD based on the MapReduce paradigm to solve the efficiency bottleneck. We evaluate SG-WSTD against several strong baselines on a real-life query log from AOL. The proposed framework demonstrates significantly improved data interpretability, better prediction performance, higher topic distinctiveness and superior scalability in the experimentation.