以跨網站使用者興趣傳遞輔助冷啟動推薦系統

本論文嘗試使用跨網站使用者興趣傳遞解決推薦系統中的冷啟動問題。一般來說，推薦系統使用評分和文字兩種資訊將感興趣的物品推薦給使用者。然而使用這兩種資訊時皆可能遇到所謂的「冷啟動問題」，也就是當使用者過往對物品的評分記錄不多，或甚至無任何記錄時，推薦系統將無法對使用者進行有效的推薦。常用於解決冷啟動問題的方法是引入輔助資料。現今許多使用者都同時活躍於多個社群網站，同時這些網站之間彼此也多有連接帳戶的機制。假設能夠從相互連接的其他網站中獲取資料，就能夠利用這些輔助資料解決新進用戶的冷啟動問題。然而上述的「跨網站興趣傳遞」策略中隱含著一個重大的難點，即不同網站之間往往並不具有相同的物品評分機制，或相同的文字資料結構。本論文的優勢在於能利用不具特定結構的一般文字，將一網站的資訊帶到另一網站的推薦系統之中。確切來說，本論文使用主題模型從一網站抽取關於使用者的特徵向量，以計算使用者之間的相似度。並修改機率型矩陣分解模型，利用相似度計算最近鄰居，將最近鄰居的隱含向量收集起來，形成一組「最近鄰居虛擬樣本」。利用這組加有權重的虛擬樣本，便得以估計冷啟動用戶隱含向量的機率分配，進而對此用戶形成推薦清單。在實驗部分使用一組現實生活中的跨網站資料集驗證此方法的有效性，和過往提出的模型相比，本論文提出的方法尤其在冷啟動問題嚴重的狀況下，取得相當程度的優勢。

關鍵字

冷啟動推薦；協同過濾；轉移學習；最近鄰居法；矩陣分解

並列摘要

In this work, we attempt to transfer user interests across websites for cold-start recommendation. Both rating-based and text-based recommender systems may suffer from the cold-start problem. One effective way to ease the cold-start problem is to introduce auxiliary data. Users nowadays hold multiple accounts across websites. If data can be obtained via the account linking mechanism, there will be an abundant supply of auxiliary data. Although this cross-site approach can be exploited to solve the cold-start problem, it is often the case that we have to deal with heterogeneous data when transferring knowledge across websites. In this work, we make use of unstructured auxiliary text to solve the cold-start problem. In particular, we extract topic vectors from source-domain text, and use the similarity scores between users to construct "nearest-neighbor pseudo data", a set of weighted (pseudo) samples which can be used to estimate the unknown parameters of the distribution over the user latent factors in the target domain. The inference process and model structure of the probabilistic matrix factorization has been modified to utilize this pseudo dataset. Improvement over previous methods, especially for the cold-start users, has been demonstrated with experiments on a real-world cross-website dataset.

並列關鍵字

Cold-start problem ； Collaborative filtering ； Transfer learning ； Neighborhood methods ； Matrix factorization

參考文獻

Y. Koren, R. Bell, and C. Volinsky, “Matrix factorization techniques for recommender systems,” Computer, vol. 42, pp. 30–37, Aug. 2009.

Z. Deng, M. Yan, J. Sang, and C. Xu, “Twitter is faster: Personalized time-aware video recommendation from twitter to youtube,” ACM Trans. Multimedia Comput. Commun. Appl., vol. 11, pp. 31:1–31:23, Jan. 2015.

H. Shan and A. Banerjee, “Generalized probabilistic matrix factorizations for collaborative filtering,” in Proceedings of the 2010 IEEE International Conference on Data Mining, ICDM ’10, (Washington, DC, USA), pp. 1025–1030, IEEE Computer Society, 2010.

J. Blömer and K. Bujna, “Simple methods for initializing the EM algorithm for gaussian mixture models,” CoRR, vol. abs/1312.5946, 2013.

R. Pan, Y. Zhou, B. Cao, N. N. Liu, R. M. Lukose, M. Scholz, and Q. Yang, “One-class collaborative filtering,” in Proceedings of the 8th IEEE International Conference on Data Mining (ICDM 2008), December 15-19, 2008, Pisa, Italy, pp. 502–511, 2008.

國際替代計量

以跨網站使用者興趣傳遞輔助冷啟動推薦系統

全文下載

主題瀏覽