A Real-time Embedding Increasing for Session-based Recommendation with Graph Neural Networks

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/88320

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/88320

Title:	A Real-time Embedding Increasing for Session-based Recommendation with Graph Neural Networks
Authors:	江明勳;Jiang, Ming-Shiun
Contributors:	資訊工程學系
Keywords:	機器學習;推薦系統;圖神經網路;machine learning;recommendation system;Graph Neural Networks;session;unknown item
Date:	2022-06-29
Issue Date:	2022-07-13 22:46:28 (UTC+8)
Publisher:	國立中央大學
Abstract:	隨著機器學習研究的不斷進步，在沒有大量複雜數據的情況下獲得良好的性能，有時比要求模型從大量數據中獲得良好的性能更重要。在推薦系統領域，用有限的數據挖掘用戶的興趣是熱門的研究方向之一。基於會話圖神經網路是一種非常流行的推薦模型，它只需要簡單的用戶瀏覽記錄就可以做出很好的推薦，但是這種模型通常有一個明顯的缺點，它不能對模型在訓練階段沒有見過的未知項目執行任何操作，就算它不是冷啟動項目也一樣。這在實際應用中是一個大問題，機器不太可能重複訓練大型模型，會消耗大量資源。為了解決這個問題，本文提出了一種新穎的可控式添加方法，可以在不影響原始性能的情況下盡可能地添加有用的表示。在許多真實世界數據集上進行的大量實驗表明了我們方法的有效性和靈活性，並且它也有機會和潛力用於其他模型或其他任務。;As the research of machine learning continues to progress, achieving good performance without a large amount of complicated data is prioritized over asking the model to reach a good performance from huge data. In the field of recommendation systems, digging out users′ interests with limited data is one of the popular research directions. Session-based recommendations with Graph Neural Networks is a very trendy model, it can make a good recommendation with only simple user browsing records, however, this kind of model usually has an obvious disadvantage, it can not perform any actions on an unknown item which model have not seen during the training phase, even though it is not a cold start item. This is a big problem in practical applications, machines are unlikely to train the large model repeatly, at it will consume a lot of resources. To solve this problem, a novel controllable addition method is proposed, the useful representations can be added without affecting the original performance as much as possible. Extensive experiments conducted on many real-world datasets show the effectiveness and flexibility of our method, and it also has the opportunity and potential to be used in other models or other tasks.
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	55	View/Open

社群 sharing

Loading...