針對時間有限的產品之以排名為基礎的推薦式系統

Abstract

在這份論文當中，我們解決了物品有生命時間限制的排序式推薦問題。我們考慮``時間有限''的物品可能會在某個時間點後失效。透過修改貝氏機率排序模型，這個模型能夠優化排序推薦，我們解決單一類別協同式過濾問題。 根據如此，我們提出兩個增強的模型，一、針對生命時間有限的物品，一個取負樣的技巧。二、一個可以把物品生命週期的時間特色模擬進來的模型。 透過在Kiva.org的資料實驗，我們驗證我們的增強模型有效。

In this work, we solve the ranking-based recommendation task given products with limit life in predicting the preference rank of users on items. We consider the `time-limit'' items that can be expired after certain point. We propose to modify the Bayesian Personalize Ranking (BPR) framework to solve the one-class collaborative filtering problem, which directly optimizes the task for ranking. We proposed two enhancements: A time-limit negative sampling technique attempts to solve the ranking predicting task for time-limit items; and a time-aware model attempt to model the temporal effect in its whole life. Experiments on Kiva.org datasets show the merit of the proposed enhancement.