基於地點類別轉移和時間行為的連續興趣點推薦系統

Abstract

智慧型裝置與網路的普及加上GPS系統的快速發展，使用者能夠在社群網路上利用"打卡"的動作來分享到訪某個地點的經驗。透過分析大量的打卡資料，可以找出使用者的偏好以及行為模式，而因此提供很大的商機給第三方服務發展個人化的服務，例如，興趣點 (Point-of-interest, POI) 推薦系統。興趣點推薦至今已被廣泛且深入的研究，衍伸出的連續性興趣點推薦 (Successive POI) 在近幾年由於命題較切中使用者使用情境而更被廣泛探討。但大部分的方法分析使用者連續性打卡行為都是間接考慮多項因素，而沒有分析使用者和連續性打卡行為的直接關聯。除此之外，大量稀疏的資料導致的效率問題也是興趣點推薦的挑戰之一。因此我們提出兩層式的架構，首先，利用矩陣分解 (Matrix Factorization) 去分析使用者和使用者對於興趣點類型的連續性打卡行為的直接關聯，加上時間因素對興趣點類型的影響，來預測使用者對於興趣點類型的偏好。接著，把那些偏好高的興趣點類型的興趣點留下，然後把使用者偏好、時間影響、地理上的影響等因素結合在一起，來做最後的興趣點推薦。在 Gowalla 打卡資料的實驗結果顯示，我們的推薦系統比現有的連續興趣點推薦系統效能和效率上都更好。

With the popularization of smart device and internet and the rapid extension of GPS system, people are able to share their experience on locations in social network through "check-ins". By analysing the huge amount of check-in data, users' preferences and their behavior patterns can be investigated and hence provide a good opportunity for third party service to develop personalized service, such as, POI recommendation. POI recommendation has been widely researched these days, but successive POI recommendation which is extended from POI recommendation attracts extensive attention nowadays because the problem meets the real needs. However, most of previous works analysed users' sequential check-in behavior by indirectly considering multiple factors rather than considering the interaction between users and their sequential check-in behavior directly. Besides, the low efficiency caused by the huge amount of sparse check-in data is another challenge in POI recommendation. Therefore, we propose a two-step approach to build our recommendation system. First, we utilize Matrix Factorization technique to analyse the interaction of users and their sequential check-in behavior on POI categories directly. Combined with the impact of temporal influence on POI categories, a category list that users have more interested in is predicted. Then, after removing those POIs not in the predicted category list, we fuse user preferences, temporal influence and geographical influence into a unified POI recommendation system and finally recommend POIs to users. The experimental result on Gowalla check-in dataset shows our recommendation system is better than several state-of-the-art methods both on effectiveness and efficiency.