On the Semantic Annotation of Places in Location-Based Social Networks Mao Ye1, Dong Shou1, Wang-Chien Lee1, Peifeng Yin1, Krzysztof Janowicz2 1Department 2Department of Computer Science and Engineering The Pennsylvania State University {mxy177,dus212,wlee,pzy102}@cse.psu.edu of Geography University of California, Santa Barbara {jano}@geog.ucsb.edu Introduction and Motivation Location-based Social Networks Tags are missing Tags are important E.g., Facebook Place and Foursquare In our Foursquare and Whirrl dataset, there are a lot of places missing tags Business categorization Location search Place recommendation Places with tags Data cleaning … 67% User check-in places Places missing 33% tags Places with tags Places missing 32%tags 68% Foursquare Problem Description Whrrl SAP Framework Place semantic annotation (SAP) problem Check-in logs Binary Classifier For tag t1 Multi-label classification problem Input Place User check-in logs <who, where, when> Binary Classifier For tag t2 Feature Extraction (FE) Component Feature extraction (FE) Some places are tagged Binary Classifier For tag tm Check-in logs features Output Features to describe a place Infer tags for the rest places FE- Explicit Pattern FE- Implicit Relatedness Day 1 Day 2 00:00 Day 3 Restaurant Restaurant Spa Shopping Restaurant 23:59 Total number of unique visitors Maximum number of check-in of a single user Restaurant Day 6 Day 7 Day 8 Bars Health ? Restaurant Total number of check-in Day 5 Bars Gym EP Feature List Day 4 Beauty Restaurant Restaurant Restaurant Shopping Shopping Restaurant Restaurant Bars Places checked in by the same user at around the same time (not necessarily the same day) are probably in the same category Daily probability of check-in Hourly probability of check-in Evaluation Whrrl.com. 5,892 users, 53,432 places and 199 types of tags Comparison: EP, IR and SAP (EP+IR) Category Restaurant&Food (Res) Shopping (Sh) Nightlife (NL) Percentage 37% 18% 19% According to Yelp, we map 199 tags into 21 categories.