以使用者偏好於不同時間區間的整合來改善電子商務推薦系統

Abstract

在訓練電子商務的推薦系統模型時,使用者和商品之間的互動是資料的核心,例如統計過去一段時間內,使用者對商品的購買或觀看次數,但是這個次數不包含時間資訊,為了解決此問題,我們提出以時間權重調整互動次數的方法,這個方法是一個框架,稱 User Preference Ensemble with Time Intervals (UPETI),讓推薦模型 能把時間權重考慮進使用者與商品的互動,藉由整合不同時間區間的使用者偏好度,得到最終的使用者偏好度。在我們的實驗中,使用創業家兄弟的好吃市集網站的資料,其中有 3 萬多位使用者和 5 千多件商品所形成的 51 萬多筆互動紀錄。結果顯示,套用 UPETI 框架的 LightFM 和 WMF 分別被改善了 6.20%和 4.20%,表示 UPETI 框架是可行的。除此之外,UPETI 框架極有彈性,可以套用到任何現有模型且不用做任何更改。

Interactions between users and products are the core of data for modeling electronic commerce recommendation systems. For example, we can count the number of “buys” or “clicks” in a certain period between a user and a product. However, the interaction usually does not have time stamp associated with it. To handle this problem, we propose a way to adjust the number of interactions with time weights. In particular, we propose a framework called UPETI (User Preference Ensemble with Time Intervals) for any recommendation model to take time weights of interactions into consideration, and the final user preference by integrating the user preferences from different time intervals. In our experiments, we used the dataset from Food123 of Kuobrothers with thirty thousand active users or so and five thousand active products or so, with around fifty-one thousand interaction records. Experiment results show that with UPETI framework, LightFM and WMF can achieve relative improvement of 6.20% and 4.20%, respectively, indicating the feasibility of the proposed framework. Moreover, UPETI framework is flexible and it can be applied to any existing models with also no modification.