Dissertation/ Thesis
Session-Aware Recommendation Using Hierarchical Recurrent Network
| العنوان: | Session-Aware Recommendation Using Hierarchical Recurrent Network |
|---|---|
| Alternate Title: | 以階層式遞歸網路建構之階段感知推薦 |
| المؤلفون: | CHEN, PO-YU, 陳柏宇 |
| Thesis Advisors: | LIN, MING-YEN, 林明言 |
| سنة النشر: | 2019 |
| المجموعة: | National Digital Library of Theses and Dissertations in Taiwan |
| الوصف: | 107 Recommender system is one of the core elements in modern business applications. Traditional recommendations such as content-based or collaborative filtering usually require users’ profiles to establish the recommendations. Session-based recommendations recently attract researchers’ interests because the recommendation can be obtained from recent user interactions without user background information. Current session-based recommendations generally use only the last session for model constructions. Nevertheless, ignoring the historical sessions of a user’s interaction makes the recommendations less accurate, especially in electronic commerce (EC) applications. Considering all the sessions of a user’s interaction, session-aware recommendation may further increase the accuracy of recommendations. Recurrent neural networks (RNN) are frequently adopted to establish the recommendations since sessions are temporal items in session-aware recommendations. However, not all the past sessions are effectively engaged in recent RNN methods such as II-RNN recommendations. Therefore, we propose the Weighted II-RNN method to greatly increase the recommendation accuracy. Weighted II-RNN is a two-layered architecture, the outer RNN learns historical sessions and generates temporal characteristics for the learning of the intra-RNN, which layer generates an average characteristic during learning. Our experiments using two popular EC datasets show that, in average, Weighted II-RNN outperforms Most-popular and Item-kNN by increasing the recall for 80% and the MRR for 103%. Weighted II-RNN also outperforms RNN and II-RNN, by increasing the recall for 50% and the MRR for 148%. |
| Original Identifier: | 107FCU00392023 |
| نوع الوثيقة: | 學位論文 ; thesis |
| وصف الملف: | 57 |
| الاتاحة: | http://ndltd.ncl.edu.tw/handle/4n3vr5 |
| رقم الانضمام: | edsndl.TW.107FCU00392023 |
| قاعدة البيانات: | Networked Digital Library of Theses & Dissertations |
| الوصف غير متاح. |