採用專家決策軌跡之深度學習推薦系統

Abstract

隨著網路科技的發展，人已無法負擔每日所接收的資訊量。推薦系統的出現讓人能夠有效率地搜尋與獲取資訊，而目前主流的推薦系統包含三大類：內容導向、協同過濾、混合型。我們發現這些推薦系統皆沒有考慮到項目間的順序關係，而在我們日常生活中搜尋資訊、挑選商品等，卻是經常是有順序性的。 因此本研究提出一個以循環神經網路及專家權重為核心的推薦系統EXTRA，透過匯集使用者們的序列預測結果，加上專家權重的調整，產生出準確的推薦清單。使用者的序列預測模型，是透過使用者與物件互動的歷史紀錄來訓練，預測出下一個使用者會想要互動的物件。專家權重則是透過使用者本身的資料及其互動過的物件，計算出使用者的權重，權重越高即代表該使用者的序列預測結果影響力越大。 本研究以臺灣知名線上論壇Mobile01與PTT的資料來實作與評估EXTRA。從實驗結果我們可以確定在論壇討論區推薦的問題上，EXTRA的表現遠比內容導向、協同過濾等方法來得好，也證明了EXTRA是有辦法適應不同平台及不同討論主題。此外，我們還發現添加了專家權重確實提升了EXTRA的準確率。 本研究提出的方法提供了推薦系統領域一個發展的可能性，考慮物件間的序列關係及加入專家權重都是可以有效提升推薦準確率。本研究提供一個簡單的概念模型，透過與現在蓬勃發展的機器學習領域結合，不論是在序列預測模型上，還是在語意分析上，EXTRA也許還有被改進的空間。此外，EXTRA並非只可應用在論壇討論區推薦的問題上，也可嘗試應用在其他場域，還有待後續研究再對EXTRA進行更進一步的實驗與分析。

With the development of Internet, people cannot process the enormous amount of information received every day. The recommender system enables people to search and obtain information efficiently. The mainstream recommendation methods include con-tent-based method, collaborative filtering method, and hybrid method. However, we find that these methods do not consider the sequential relations between items. In our daily life, searching for information or selecting goods often follows a specific order. In this study, we propose a recommender system, EXTRA, with a recurrent neural network and expert weights as its core. By aggregating the users’ predicted viewing se-quences adjusted further by expert weights, an accurate recommendation list can be generated. The trace prediction model for users is trained by the historical records of the users’ interactions with items, which is able to predict the item with which a user would want to interact next. The expert weight for a user is obtained from users’ own infor-mation and items interacted by the user. The higher the expert weight, the greater the influence of the user’s predicted viewing sequences. To implement and evaluate EXTRA, we use the data collected from Mobile01 and PTT, two most popular online forums in Taiwan. From the experimental results, we can confirm that the performance of EXTRA is far better than that of content-based method, collaborative methods, etc. Moreover, we proved that EXTRA can adapt to different platforms and different topics. In addition, we also find that adding expert weights does improve the accuracy of recommendation list provided by EXTRA. The method proposed in this study provides a new direction in the field of recom-mender systems. Considering the sequential relations between items and adding expert weights can effectively improve the performance of a recommender system. In this re-search, we provide a conceptual model which may be improved further by combining other machine learning and semantic analysis techniques. Moreover, EXTRA is not only applicable to forums, but can also be applied to other fields. Further experiment and analysis to EXTRA can be done in the future.