序列推薦模型之轉換學習

Abstract

本論文的研究是將轉換學習結合到序列推薦模型。目前最先進的推薦模型可以學習使用者的喜好，並據此對不同的使用者產生不同的推薦。然而，對於缺乏資料的使用者，個人化的推薦系統無法學習到他們的偏好，也就無法針對物品有精準的排序。近來，為了解決這個問題，轉換學習被應用到推薦系統上，雖然新環境可能因缺乏資料而導致模型擬合不足，但可以利用舊環境的資料來幫助模型訓練。然而，大部分結合轉換學習的推薦系統，都是針對評分的問題，這種問題的使用者回饋都是外顯的，也不會有順序的關係。在本篇論文，我們將轉換學習應用到序列推薦模型上。我們會用現實世界的資料來進行實驗，並據此展示我們提出框架的有效性。

In this work, we attempt to apply transfer learning to sequential recommendation model. Most of the state-of-the-art recommendation systems consider user preference and give different results to different users. However, for those users without enough data, personalized recommendation systems cannot infer their preference well and then rank items precisely. Recently, transfer learning techniques are applied to address this problem. Although the lack of data in target domain may result in underfitting, data from auxiliary domains can be utilized to assist model training. However, most of recommendation systems combined with transfer learning aim at rating-based problems whose user feedback is explicit and not sequential. In this paper, we want to apply transfer learning techniques to a model utilizing user preference and sequential information. Experiments on real-world dataset are conducted to demonstrate the effectiveness of our framework.