透過您的圖書館登入
IP:18.234.232.228
  • 學位論文

行為關聯項目為基礎之協同過濾推薦系統

A Behavior Related Item-based Collaborative Filtering Recommendation System

指導教授 : 段裘慶

摘要


近年來,隨著無線網路與行動計算技術日益成熟,使用者能夠在任意時間及地點,透過行動裝置向資料庫伺服器取得相關服務與資訊,其中位置相依之地標興趣點(POI)推薦系統,所提供多元化查詢服務為最符合使用者之需求。在過去推薦系統研究領域中,多數採用協同過濾改善推薦品質,以達到資訊推薦個人化之效果,卻忽略用戶具有移動特性,若只單純考慮POI偏好評比改善其推薦品質,將導致推薦結果不符合用戶期望。然而,考量行動用戶所在區域性資訊POI項目外,偏好評比值亦是推薦用戶影響因素之一,本研究認為調整評比值有助於改善推薦品質。 為了改善傳統推薦系統未考量移動特性而造成的缺失,本研究提出以用戶行為模式為基礎之行為關聯評比推薦系統(BRR),係整合了行為關聯探勘及推薦技術應用,幫助用戶篩選出可能感興趣之熱門物件,以項目為基礎之協同過濾進行POI價值推薦。透過參考偏好程度適時調整評比,藉此達到滿足用戶推薦準確性與偏好調適之目的。 效能分析上與位置導向之個人化感興趣點推薦機制、多階段協同過濾、適地性旅館資訊推薦模型進行效能評估,並定義其推薦精確度、推薦覆蓋率、平均回應時間作為效能分析之評量指標。從模擬實驗結果顯示,本研究所提出之BRR推薦機制多數情況下,均具較高推薦精確度與較少平均回應時間;在推薦項目數為10個且用戶連續發出10次查詢行為時,BRR之推薦精確度優於其他適地性推薦機制平均約10%至51%。以推薦覆蓋率而言,相較其他策略平均提昇約9%;雖然BRR須預先處理行動用戶之行為關聯,於平均回應時間增加0.41秒,但隨著連線用戶數之遞增,平均回應時間可縮短至0.17秒。

並列摘要


In recent years, with increasingly sophisticated wireless networks and mobile computing technologies, users access relevant services and information via mobile devices from the database servers. The location-dependent landmark Point of Interest (POI) recommendation system could satisfy the user needs for its diversified query services. Previous recommendation system-related studies mostly improved the recommendation quality by collaborative filtering to achieve the effect of personalized information recommendation. To consider the mobility of users in the traditional recommendation systems, this paper proposed the Behavior-related Rating Recommendation (BRR). The proposed BRR integrated the behavior-related mining and recommendation technologies to select highly interesting objects for users and carry out POI value recommendated by content-oriented collaborative filtering. As preference rating declines over time, the timely adjustment of rating by referring to preferences may promote the accuracy and object preference in recommendation systems. This study conducted experiments to measure performance of ULPPR, MSCF, LSRM. The metrics we defined included recommendation accuracy, recommendation coverage, average response time for analyzing performance. The simulation results suggested that the proposed BRR recommendation mechanism had relatively high recommendation accuracy and shorter average response time in most cases. For 10 users’ with 10 times of queries, the BRR recommendation accuracy was better than other recommendation mechanisms by 10-51% on average. In terms of recommendation coverage, BRR improved by 9% on average as compared with other strategies. Although the BRR requires to pre-process behavior associations of the mobile users with average response time increasing by 0.41 seconds, the average response time can be reduced to 0.17 seconds with of online users.

參考文獻


[57]林彥亨,以資訊關聯度為導向之近鄰預取查詢策略,碩士論文,國立臺北科技大學電腦與通訊研究所,臺北, 2009。
[48]J. P. Huang, S. J. Chen, and H. C. Kuo, “An Efficient Incremental Mining Algorithm-QSD,” Intelligent Data Analysis, vol. 11, no.3, pp. 265-278, 2007.
[1]B. Bruegge and B. Bennington, “Applications of Wireless Research to Real Industrial Problems. Applications of Mobile Computing and Communication,” IEEE Personal Communications, vol. 3, no. 1, pp. 64-71, Feb. 1996.
[3]J. E. Padgett, C. G. Gunther, and T. Hattori, “Overview of Wireless Personal Communications,” IEEE Communications Magazine, vol. 33, pp. 28-41, Jan. 1995.
[4]H. Maass, “Location-Aware Mobile Applications Based on Directory Services,” ACM Mobile Networks and Applications, vol. 3, no. 2, pp. 157-173, Aug.1998.

延伸閱讀