Discovering Co-located Queries in Geographic Search Logs

Though general query log mining has been studied for years, little work has been done to utilize geographic search logs. A geographic search request basically contains a query consisting of one or more keywords, and a location specifying the geographic search area. In this paper, we study the problem of discovering co-located queries which are frequently searched over nearby locations. One example co-located query pattern is "shopping mall", "parking". This pattern indicates that people often search "shopping mall" and "parking" over locations close to one another. We formally define co-located query patterns and propose two approaches to mine the patterns. The basic approach is based on an existing spatial mining algorithm. To find more co-located queries that only appear in specific regions, we propose a lattice based approach. It divides the geographic space into regions and mines patterns in each region. We also define a locality measure to categorize patterns into local and global. Experimental results show that the lattice based approach achieves higher performance than the basic approach in the number of patterns, the quality of patterns, and the proportion of local patterns.