基於矩陣分解之賓客推薦

Guest Recommendation Based on Matrix Factorization

林聖庭(Sheng-Ting Lin)

指導教授 ： 戴志華

國立臺北大學/電機資訊學院/資訊工程學系/碩士(2016年)

摘要

社交活動是人們聯繫彼此感情與拓展交友重要的一環，而人們越來越仰賴社交網路平台，更多樣的社交活動需求與應用也應運而生，朋友之間舉辦聚會也就變得更加頻繁。在舉辦活動時，主辦人在自己大量的朋友中要去搜尋自己想要邀請的朋友，會是一件十分不簡單的事情，因為我們的朋友數很多，每位朋友的特性也不同，想要邀請那些朋友就會是令主辦人十分頭疼的事情。因此，我們提出了一個Guest Invitation list for Hosts (GIH) 的推薦問題，目的是在活動剛建立的時候，幫助主辦人從既定朋友中過濾出有哪些朋友是主辦人想要邀請的，最後我們會推薦一個賓客的排序清單供主辦人做參考。
在推薦系統中，我們分成兩大部分來進行，分別是訓練還有推薦的部分，在訓練的部分我們將蒐集的活動相關資料做前處理後，再利用活動的標題與描述為相同的活動做分群達到資料的穩定性的效果，也可以讓相似的活動被分在一起；我們將活動中的關係用矩陣的方式建立，並使用五種不同的矩陣分解方法來尋找出活動背後可能的邀請因素，五種演算法分別是Naïve 、NMF、Density、TFP還有NMF+TFP。而在推薦階段會因演算法的設計概念不同而有不同的算法，五種演算法中擁有最佳的效能的會是NMF+TFP，可以達到0.87的準確率。
在我們的實驗中，不只對所有的活動做總平均，也針對不同類型的活動做個別的討論，另外也包括了個別的參數設定實驗還有分群的實驗，同時也證實了邀請時的考量對於人來說一般都不會太多，但是單一因素考量的也不會太好，最後我們再拿出三個不同類型的活動作案例的探討。

關鍵字：活動、主辦人、賓客、矩陣分解、推薦

關鍵字

活動 ； 主辦人 ； 賓客 ； 矩陣分解 ； 推薦

並列摘要

Social activities are the important parts of life for connecting and developing friendship. As online social platforms such as Facebook and Twitter get popular, more and more social applications have also been developed, which greatly facilitate making new friends and organizing activities online. Nevertheless, organizing activities is still time-consuming, since a host needs to search through all his or her friends to figure out whom to invite and suffers if he or she has hundreds or even thousands of friends online. Thus, in this thesis, we address a recommend problem name Guest Invitation list for Hosts (GIH). The object of GIH problem is to rank the friends of a host under the consideration of an activity and to recommend a proper guest invitation list for the host.
For the GIH problem, we develop a system consisting of training and recommendation phases. At the training phase, we collect activity of similar types to learn the associations between the activities and invited guest. To identify reliable associations, the collected activities will first be clustered according to the titles and descriptions of activities. After that, we establish a matrix to consider the relationship between the host and invited guests of an activity, and propose five matrix factorization methods, i.e., Naïve, NMF, Density, TFP and NMF+TFP, to derive the associations from the matrix. At the recommendation phase, a new activity is first categorized to the most similar activity group according to its title and descriptions. Then, the guest recommendation is performed based on the learned associations of that activity group. Experimental results show that NMF+TFP has the best performance. We also explore the effects of various parameters, investigate the performances of simple rules and complex rules, and perform case studies for three types of activities.

Key-word(s)：Activity, Host, Guest, Matrix factorization, Recommendation

並列關鍵字

Activity ； Host ； Guest ； Matrix Factorization ； Recommendation

參考文獻

[34] V. P. PROCa, F. Shahnaz, M. W. Berry and R. J. Plemmons. "Text Mining using non-Negative Matrix Factorizations." Proceedings of the SIAM International Conference on Data Mining Society for Industrial and Applied Mathematics, 2004.

[2] The Chinese Knowledge and Information Processing (CKIP) website.

[9] The Pew Research Center website.

[12] R. Alves, D. S. R. Baena and J. S. Aguilar-Ruiz. "Gene association analysis: a survey of frequent pattern mining from gene expression data." Briefings in bioinformatics 11(2): 210-224, 2010.

[13] R. Burke. "Hybrid recommender systems: Survey and experiments." User modeling and user-adapted interaction 12(4): 331-370, 2002.