推薦系統之概率圖模型實踐排序學習探討

Abstract

在網路通訊科技的日新月異的世代，資訊過載(information overloading) 的問題普遍存在各種網路應用當中，因此電子商務領域中普遍採用推薦系統(recommender system)，幫助使用者取得高品質的資訊。推薦系統基本機制是產生個人化推薦清單，依照個別使用者的喜好順序排列物品，推薦他們可能喜歡，但尚未看過的項目，達到提升顧客體驗與收益的目的。推薦系統的過濾資訊概念分成內容(content-based)與協同(collaborative)兩大方向，協同過濾能簡易且有效的產出衡量喜好指標，並保持推薦項目的差異化，其中混合模型(mixture model)是實踐協同過濾的一種機率模型，是最具彈性及嚴密數學推導的方法。過去推薦系統一直專於準確預測項目喜好分數，近年研究皆指出排序學習(learning to rank)更貼近使用者需求，因此成為了現今推薦系統領域的主要研究方向。本篇論文將混合模型和預測效果最優越的的列表式(listwise)排序法結合，促使混合模型一舉向排序學習領域往前推進，改變傳統模型準確預測分數卻不符合推薦目的的方式，同時也善盡混合模型的參數可擴充性，解決推薦系統的最常見的遺漏值(missing theory)議題。最後在嚴謹的強推論交叉驗證實驗中，證實了模型優化混合模型的可能性，成功地讓機率模型在推薦系統領域中更加完整。

Facing the rapid growing volume of information across the internet, most of commercial websites have adopted recommender systems to help their customers get valuable information efficiently. As the goal of recommender systems is to provide a ranking list of items according to user preferences, ranking has become the core of the systems. In recent year, a novel technique called learning to rank (LTR) which resolves ranking problem with machine learning algorithms has attracted researchers of recommender systems. In this paper, we investigate learning to rank and recommender systems. Specifically, we incorporate the listwise learning to rank approach into a mixture model(MM) to learn the preferences of users in a recommender system. The listwise LTR approach has been shown significantly better than the pointwise and pairwise approaches because it takes a whole ranking list of items as a learning instance that matches the goal of recommender systems. Also, the MM is effective in modelling user preferences that enhance collaborative filtering (CF) to identify the similar taste users of a target user. In addition to incorporating LTR into recommender systems, we investigate non-random missing data, and introduce effective model parameters to capture the missing mechanism into the MM. Our main contribution is presenting listwise mixture model (list-MM) perfectly incorporate CF, LTR, and MM. This model is processed of accuracy, flexibility, and scalability which is suitable for modern recommender systems.