騰訊微博是中國最大的微博網站之一，騰訊微博上超過200萬的註冊用戶，每天產生超過40萬筆的消息。為了避免使用者處於資訊過載的狀況，因而，騰訊微博推薦使用者可能會感到有趣的項目。本篇論文為預測使用者是否會點擊跟隨騰訊微博所推薦的項目。論文分為兩大部分：第一部分為判斷使用者的喜好，第二部分則是判斷使用者是否專注於推薦項目上。判斷使用者喜好的部分，我們建立了多種Model based Collaborative Filtering模型以及Content based的模型來模擬使用者的喜好。第二部分則以資料的時間序列來建立Occupied model以模擬使用者處於何種狀態。最後，合併Occupied model與使用者喜好模型為最後的預測模型。在本篇論文我們以Session為單位來計算模型的Hamming loss，使用者喜好模型與Occupied model合併後的Hamming loss都會明顯下降，並達到最低的Hamming loss 0.13。

Tencent Weibo is one of the largest micro-blogging websites in China. There are more than 200 million registered users on Tencent Weibo, generating over 40 million messages each day. Recommending appealing items to users is a mechanism to reduce the risk of information overload. The task of this paper is to predict whether or not a user will follow an item that has been recommended to the user by Tencent Weibo. The paper contains two parts: predicting users’ interests and distinguish whether the user is busy or available to browse recommended items. We apply several models based collaborative filtering as well as content-based filtering to capture users’ interests. Besides, we built an occupied model to distinguish users’ state and combined with recommendations methods as the final result. In the paper, we used session-based hamming loss as performance measure. The hamming loss of recommendation methods were greatly reduced (40%) above with occupied model from 0.187 to 0.13.

[1] C. D. Meyer, ‘‘Matrix Analysis and Applied Linear Algebra”, SIAM, p. 514, 2000.
[2] D. Goldberg, D. Nichols, B. M. Oki, and D. Terry, ‘‘Using Collaborative filtering to weave an information tapestry”, Communications of the ACM, Volume 35, Issue 12, pp61-70, 1992.
[3] D. D. Lee, H. S. Seung ‘‘Algorithms for non-negative matrix factorization”, In: Advances in Neural Information Processing Systems. 2000. 556−562, 2000.
[4] FICO, 2012, FICO Model Builder 7. http://www.fico.com/en/Products/DMTools/Pages/FICO-Model-Builder.aspx
[5] G. Golub and K. Kahan ‘‘Calculating the singular values and pseudo-inverse of a matrix.” Journal of the Society for Industrial and Applied Mathematics, 1965,2(2):205−224. [doi: 10.1137/0702016], 1965.
[6] G. Tsoumakas and I. Katakis, ‘‘Multi-Label Classification: An Overview”, International Journal of Data Warehousing & Mining, 3(3), 1-13, 2007.
[7] I. S. Dhillon, S. Mallela, and D. S. Modha, ‘‘Information-Theoretic Co-clustering”, Proceedings of The Ninth ACM SIGKDD Conference on Knowledge Discovery and Data Mining(KDD), pages 89-98, 2003.
[8] K. Kiwiel, ‘‘Convergence of Approximate and Incremental Subgradient Methods for Convex Optimization”, SIAM Journal on Optimization (SIAMJO) 14(3):807-840, 2004.
[9] L. Liu and M. Tamer, ‘‘Mean Average Precision’’, Encyclopedia of Database Systems2009: 1703, 2009.
[10] LibFM: Factorization Machine Library. http://www.libfm.org/
[11] P. Melville and V. Sindhwani, ‘‘Recommender Systems”, Encyclopedia of Machine Learning 2010:829-838, 2010.
[12] R. Salakhutdinov and A. Mnih ‘‘Probabilistic matrix factorization.”, In: Advances in Neural Information Processing Systems. 2008,20(3): 451−432, 2008.
[13] S. Rendle, ‘‘Factorization machines with libFM”, ACM TIST 3(3):57, 2012.
[14] S. Rendle,‘‘Social Network and Click-through Prediction with Factorization Machines”, KDD-Cup Workshop, 2012.
[15] Tianqi Chen of team ACMClass@SJTU, ‘‘Combining Factorization Model and Additive Forest for Collaborative Followee Recommendation”, KDD-Cup Workshop, 2012.
[16] X. Zhao of team FICO Model Builder, ‘‘Scorecard with Latent Factor Models for User Follow Prediction Problem”, KDD-Cup Workshop, 2012.
[17] Y. Kluger, R. Basri, J. T. Chang and M. B. Gerstein, ‘‘Spectral Biclustering of Microarray Data: Coclustering Genes and Conditions”. Genome Research 13 (4): 703–716, 2003.
[18] Y. Koren, R. Bell, and C. Volinsky, ‘‘Matrix factorization techniques for recommender systems”, IEEE Computer 42(8):30-37, 2009.
[19] Y. Niu, Y. Wang, G. Sun, A. Yue, B. Dalessandro, C. Perlich, and B. Hamner, ‘‘The Tencent Dataset and KDD-Cup’12” KDD-Cup Workshop, 2012.