透過您的圖書館登入
IP:3.143.239.50
  • 學位論文

基於端對端軟性分群之推薦系統

Group-­Assisted Recommendation via End­-to-­End Soft Clustering

指導教授 : 鄭卜壬

摘要


有鑒於近年來電商與個人化服務的興盛,基於隱性反饋的個人化推薦扮演了不可或缺的角色,從過去的矩陣分解模型,到近年來的圖卷積網路都取得了一定的成效。但是,由於隱性反饋本身的特性,我們只能看出使用者與物品之間有無互動關係,而非其真正的偏好程度,這造成了資訊的稀缺性。另一方面,經社會科學研究發現,不同群體的人在消費與決策上有著很大的差異。雖然在現今的電腦科學領域,已經有許多以分群來輔助推薦的研究,但這些方法通常需要如商品資訊、使用者社群網路、評論紀錄等額外資源,導致其在實務上的限制。 在這篇論文中,我們提出了一個基於「端對端學習」及「軟性分群」技術的模型,在不使用額外資源的情況下,自動的去學習使用者與物品的分群,並探究群與群之間的關係。我們針對模型的結構,提出了多種變形,並比較之間的效力差異。實驗證明我們所提出的方法在兩個密度不同的真實資料集中,都有統計上高度顯著的進步,並能在不影響高活躍度使用者的情況下,提升對低活躍度使用者的推薦品質。

並列摘要


Personalized recommendation based on implicit feedback plays a vital role in our daily lives. Due to the characteristic of implicit feedback, we can only observe whether a user interacted with the item, which results in information sparsity. In addition, Social Science researchers found that different groups of people have an enormous difference in consumption and decision-making. However, existing works that assist recommendation via clustering rely on extra information, like item context and social features. In this paper, we propose a model that incorporates end-to-end learning and soft clustering. It learns the group assignment and embedding simultaneously without using extra resources. We conduct extensive experiments to evaluate the effectiveness and reasonableness of our proposed model. Experimental results show that our approach improves with a significant level on two real-world datasets.

參考文獻


Yehuda Koren, Robert Bell, and Chris Volinsky. Matrix factorization techniques for recommender systems. Computer, 42(8):30–37, 2009.
Brian Wansink, Matthew M Cheney, and Nina Chan. Exploring comfort food pref­erences across age and gender. Physiology behavior, 79(4­-5):739–747, 2003.
AL Reicks, JC Brooks, AJ Garmyn, LD Thompson, CL Lyford, and MF Miller. Demographics and beef preferences affect consumer motivation for purchasing fresh beef steaks and roasts. Meat Science, 87(4):403–411, 2011.
Anthony Palomba. Consumer personality and lifestyles at the box office and beyond: How demographics, lifestyles and personalities predict movie consumption. Journal of Retailing and Consumer Services, 55:102083, 2020.
Gideon Dror, Noam Koenigstein, Yehuda Koren, and Markus Weimer. The yahoo! music dataset and kdd­-cup'11. In Proceedings of KDD Cup 2011, pages 3–18. PMLR, 2012.

延伸閱讀