[4O1-GS-4] Web intelligence: information recommendation / retrieval

[4O1-GS-4-04] Hybrid POI Search Combining Geographical Features and Deep Metric Learning-based Semantic Similarity

〇Yusuke Okimoto¹, Sora Takashima², Kenta Kanamori¹, Xinhu Lan¹, Junji Saikawa¹ (1. Yahoo! Japan, 2. Tokyo Institute of Technology)

Keywords:POI Search, Neural Information Retrieval, Deep Metric Learning

POI search is becoming increasingly important with the proliferation of smartphones. In recent years, natural language processing techniques based on deep neural networks (DNN) have achieved good performances in document retrieval. On the other hand, numerical information such as geographic distance and user ratings are also important features in POI search. In this study, we propose a POI search method using both semantic similarities obtained by DNN and numerical information. The semantic similarities between a query and a POI are calculated using DNN trained by metric learning. The obtained semantic similarities and numerical information are then used as features for GBDT to rank POIs. We conducted offline evaluations on the proposed method using search logs in Yahoo! JAPAN MAP and found improvement in ranking metrics.