標題: 以網格為基礎的連續興趣點推薦系統
Grid-based Successive Point-of-interest Recommendation
作者: 高洪詣
黃俊龍
Gau, Hung-Yi
Huang, Jiun-Long
資訊科學與工程研究所
關鍵字: 興趣點;推薦系統;適地性社群網路;打卡資料;網格;個人化網頁排名;Point-of-interest;Recommendation System;Location-based social networks;Check-in data;Grid;Personalized Pagerank
公開日期: 2016
摘要: 隨著攜帶式裝置的普及與發展,人們的生活習慣也逐漸在改變,使用者在社群網路上利用打卡來分享到訪某個地點的經驗以及相關資訊已成為生活中的一部分。隨著社群網路在打卡資料的累積,興趣點 (Point-of-interest, POI) 推薦系統也因應而生。透過分析過去的打卡偏好和行為來幫助使用者探索週遭,能夠為使用者提供更好的到訪體驗。而現今興趣點推薦已被廣泛且深入的研究,但針對使用者目前位置來進行推薦的方法仍有改善的空間,雖然考量興趣點間距離遠近的影響,卻忽略了興趣點的區域性質。因此我們提出一套以網格為基礎的連續興趣點的推薦系統,透過經緯度將這些興趣點切成各個方塊狀的網格來近似區域性的概念,進而評估使用者的區域性偏好。除此之外,我們將 Edge-weighted Personalized PageRank (EdgePPR) 應用到興趣點間的連續打卡轉移關係中,由使用者偏好、區域偏好、連續打卡偏好來構成我們的推薦系統。我們的資料來源為 Gowalla 和 Brightkite,為著名的適地性社群網路 (Location-based social networks, LBSNs) 服務。實驗結果顯示我們的推薦系統比現有的連續興趣點推薦系統更加準確。
With the rapid development of mobile devices, living habits of people gradually change.
Users are able to share their visiting experience to point-of-intrests (POI) on location-based social networks (LBSNs) by check-ins, which becomes a part of people's life.
With abundance of check-in data, POI recommendation service attracts a lot of interest. By analysing history check-in preferences and behaviors of users, we can not only help users to explore surroundings but also provide better visiting experience.
Nowadays, POI recommendation has been widely researched.
However, the recommendation methods that consider the current location of users, successive POI recommendation, still have room for improvement. Although the recent methods consider the geographical influence by measuring the distance, they ignore the regional influence of POI. Therefore, we propose a grid-based successive POI recommendation system, UGSE-LR.
To evaluate the region preferences of users, UGSE-LR splits geographic area into square-shaped grids for approximating the concept of regions.
Besides, UGSE-LR applies Edge-weighted personalized PageRank (EdgePPR) for modeling the sequential transitions between POIs.
Finally, UGSE-LR fuses user preferences, region preferences and sequential preferences into a unified recommendation framework.
Results on two real-world LBSNs datasets, Gowalla and Brightkite, show that our method are more accurate than the state-of-the-art successive POI recommendation methods.
URI: http://etd.lib.nctu.edu.tw/cdrfb3/record/nctu/#GT070356034
http://hdl.handle.net/11536/139899
Appears in Collections:Thesis