請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/91221
標題: | 基於預訓練的生成模型之個人化關鍵字推薦 Towards Personalized Keyword Recommendation with Pretrained Generative Models |
作者: | 陳偉倫 Wei-Lun Chen |
指導教授: | 鄭卜壬 Pu-Jen Cheng |
關鍵字: | 關鍵字推薦,預訓練,關鍵字序列編碼器,物品關鍵字解碼器, Keyword Recommendation,Pretraining,Keyword Sequence Encoder,Item Keyword Decoder, |
出版年 : | 2023 |
學位: | 碩士 |
摘要: | 近年來,序列推薦 (sequential recommendation) 的技術取得了重大進步,在捕獲複雜的用戶行為和預測序列中的下一項物品方面取得了廣泛的成功。然而,在考慮物品嵌入 (item embedding) 時面臨了一項常見的挑戰,尤其是在處理缺乏足夠用戶與物品的交互紀錄和可靠的物品嵌入的冷啟始物品 (cold-start item) 時。此外,頻繁添加新物品也對推薦系統資源密集型的再訓練產生了顧慮。為了應對這些挑戰,在本文中,我們提出通過預測物品關鍵字而非直接預測新物品的創新方法。利用用戶與物品的交互紀錄和用戶提供的關鍵字並結合無監督式的預訓練,我們的個人化關鍵字推薦模型可以捕獲用戶偏好並為下一個潛在物品推薦相關關鍵字。這種方法使得用戶能夠利用推薦的關鍵字來查找冷啟始物品的更多相關資訊,從而減輕與訓練不可靠物品嵌入相關的問題。我們在真實資料集上進行大量實驗,並驗證了我們模型的有效性,超越了幾種基準方法,並將其確立為個人
化物品關鍵字推薦之最先進解決方法。我們的貢獻有引入個人化關鍵字推薦的新問題,提高了用戶提供關鍵字的推薦準確性,提出了一種帶有關鍵字序列編碼器(keyword sequence encoder) 和物品關鍵字解碼器 (item keyword decoder) 的新模型,並展示了無監督預訓練在提高關鍵字推薦表現的好處。通過解決與冷啟始項目相關的挑戰,我們的研究開創了在不同應用場景中為用戶提供更定制、更有價值的推薦體驗之道路。 Sequential recommendation techniques have witnessed significant advancements in recent years, leading to their widespread success in capturing intricate user behaviors and predicting subsequent items in sequences. However, a common challenge arises concerning the treatment of item embeddings, particularly when dealing with cold-start items that lack sufficient interaction history and reliable embeddings. Additionally, the frequent addition of new items raises concerns about resource-intensive retraining of recommendation systems. In this paper, we propose an innovative solution to address these challenges by predicting item keywords instead of directly predicting new items. Based on user's item interactions, along with user-given keywords and incorporating unsupervised pretraining, our personalized keyword recommendation model captures user preferences and recommends relevant keywords for the next potential item. This approach enables users to utilize the recommended keywords for finding more relevant information about cold-start items, thus mitigating issues related to training unreliable embeddings. Extensive experiments on real-world datasets demonstrate the effectiveness of our model, surpassing several baseline methods and establishing it as a state-of-the-art solution for personalized item keyword recommendation. Our contributions introduce personalized keyword recommendation as a novel problem, enhance the recommendation accuracy with user-given keywords, propose a novel model with keyword sequence encoders and an item keyword decoder, and showcase the benefits of unsupervised pretraining in improving keyword recommendation performance. By addressing the challenges associated with cold-start items, our research paves the way for more tailored and valuable recommendation experiences for users in diverse application scenarios. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/91221 |
DOI: | 10.6342/NTU202304409 |
全文授權: | 同意授權(限校園內公開) |
顯示於系所單位: | 資訊工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-112-1.pdf 目前未授權公開取用 | 549.92 kB | Adobe PDF | 檢視/開啟 |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。