以時序矩陣分解方法追蹤個人喜好之概念飄移

Abstract

在這個巨量資料的時代，人們無時無刻從網際網路中接收大量的訊息。為了解決資訊過載所造成的負擔並讓使用者快速找到理想的商品，提供精準且適時的推薦對於服務供應商及使用者雙方而言皆是一個重要的議題。在推薦系統的研究領域中，評分預測是最根本的基礎問題之一，其中矩陣分解已是個被廣泛採用的方法。基於歷史的評分數據，矩陣分解利用潛在因素模型以獲得固定的使用者偏好及商品特性，然而使用者的偏好並非一成不變，反而會受到現實生活中各種因素的影響而有所變動，為了滿足其當前的品味及需求，模擬使用者偏好隨時間的演變是設計推薦系統的重要關鍵。有鑑於此，我們提出時序矩陣分解方法以追蹤各個使用者偏好之概念飄移：藉由進一步延伸stochastic gradient descent 方法及利用Lasso regression，我們針對每個使用者的心境轉變進行模擬。各方面的實驗結果皆顯示我們所提出的時序矩陣分解方法能夠 (1) 追蹤使用者偏好的演變、(2) 探討不同資料集之中使用者的心境轉變特性，以及 (3) 與原本的矩陣分解方式相比，我們所提出的方法能達到更準確的預測。

In the era of big data, people are overloaded with massive amounts of information from the Internet. As such, whether service providers can provide accurate and timely recommendations for users to quickly locate desirable items is critical to both users and service providers. In the research area of recommender systems, rating prediction is one of the most fundamental problems and the matrix factorization (MF) approach has been widely adopted for solving the rating prediction problem. The MF approach utilizes the latent factor model to obtain static user preferences and item characteristics based on the historical rating data. However, user preferences are not static but full of dynamics in the real world and therefore modeling the temporal evolution of user preferences is a key for recommender systems to satisfy users’ current taste and need. In view of this, we develop Temporal Matrix Factorization (TMF) that is capable of tracking concept drift in individual user preferences. This is done by modifying the stochastic gradient descent method for MF and modeling the transition at the individual level via the Lasso regression. Various experiments on both synthetic and several real datasets show that our TMF approach is able to (i) track the evolution of user preferences, (ii) investigate the intrinsic properties of the transition on different datasets and (iii) achieve more accurate predictions than the original MF approach.