融合多模型排序之點擊預測模型;Learning to ensemble ranking model for sequential click prediction

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/83949

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/83949

Title:	融合多模型排序之點擊預測模型;Learning to ensemble ranking model for sequential click prediction
Authors:	賴品菁;Lai, Pin-Ching
Contributors:	資訊工程學系
Keywords:	Top-N排序;推薦系統;點擊率預測;機器學習;深度學習
Date:	2020-07-20
Issue Date:	2020-09-02 17:44:41 (UTC+8)
Publisher:	國立中央大學
Abstract:	現今因為網路的蓬勃發展，網路與人們的生活密不可分，在網站的需求上，不管是購物網站或是音樂、影片網站，越來越多網站有個人化的服務及推薦。若能有效的推薦商品給使用者，不但能增加使用者對此網站的滿意度，還能讓這個網站的創造者獲得收益，造就雙贏的局面。本篇論文利用兩種不同大小的電商資料，對使用者下一個可能會點擊的商品做排序，推薦給使用者。為了將真正被點擊過的商品排在更前面的推薦序列中，我們使用了四種不同類型的模型來做排序。由於最近深度學習在各個領域表現卓越，推薦系統也開始使用深度學習架構來訓練，但我們發現，在不同的電商規模下，使用深度學習架構不一定會表現得比淺層的模型好。在發現了這個現象之後，我們使用這四個不同的模型排序過後的序列排名，來當作我們的訓練資料，並重新排序，最後我們得到了更符合使用者歷史點擊的排名，將真正被點擊過的商品的排名往前排，並獲得明顯的進步。不管在哪個資料集上，都表現的比四個基準模型還要好，並在實驗過程中發現了商品排名與商品出現次數的相關性。;With the vigorous development of the Internet, it becomes inseparable from people’s lives. Various websites, such as online retailers and online music/video streaming websites, provide personalized recommendations to show the RIGHT products to users. These recommendations may increase users’ satisfaction with these services and create revenues for the service providers. This paper aims to predict a users’ next clicking item based on two e-commerce datasets. We compared four ranking models and claimed a model performs better if the model ranks the next clicking item at a former position among the four models. Despite recent researches showed excellent performance on deep learning-based ranking models, we found that this is not always the case. Notably, we may need to consider the size of the service providers as a reference to decide to apply a deep or a shallow learning model. Since different models may work under various scenarios, we developed an ensemble model that learns to rank the items based on the rankings returned by the four ranking models. Experimental results show that the new model outperforms the four baselines: on both datasets, the new model tends to put the next clicking item at a former position.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	84	View/Open

社群 sharing

Loading...