Please use this identifier to cite or link to this item:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/88401| Title: | 透過識別出序列級上的監督式模板解決序列式推薦系統的問題 Supervised Pattern Recognition for Sequential Recommendation |
| Authors: | 劉燕芬 Yan-Fen Liu |
| Advisor: | 林守德 Shou-De Lin |
| Keyword: | 對比式學習,序列式推薦系統,自監督式學習, Contrastive Learning,Sequential Recommendation,Self-Supervised Learning, |
| Publication Year : | 2023 |
| Degree: | 碩士 |
| Abstract: | 序列式推薦系統 (SR) 藉由考慮連續性的歷史行為,來捕捉使用者瀏覽行為上的興趣動向,使下一個推薦物品更能迎合用戶的需求。 在近期序列式推薦系統的發展裡,有一分支是將對比式學習與自監督信號結合起來,以解決數據集上的資料稀疏性所造成的推薦問題;而這些方法經常透過微調數據結構的方式,來構建出略為不同的數據形態。 然而,由於這些微調的手段將對瀏覽行為中的原始意圖造成隨機擾動,使得在建構出的數據與原始資料之間,兩者反映出的語義是否仍保有一致上備受挑戰。 因此,有別於使用微調數據結構來生成自監督信號的方法,近年來的提案傾向在原始數據中,定義一個局部監督標籤(例如目標項目)來緩解對比式訊號在語意一致性上的疑慮,然而,他們所提出的局部監督式標籤,恐怕並不足以反映整個瀏覽行為序列所代表的語意。 因此,為了能進一步改善上述的問題,我們在此提出一種新穎的對比式學習框架,用意在「透過識別出序列級上的監督式模板,解決序列式推薦系統的問題」。 與先前研究有所不同的地方是,我們旨在,通過從原始數據中所挖掘出序列級的監督式模板當作對比信號,來保持其與對比目標物之間的語義能在序列級上有著一致性。 此外,我們還提出了新的對比正則式,以提高對比式學習任務在假負樣本和真負樣本的辨識能力。 我們的模型在五個公共數據集上進行的大量實驗,其結果表明我們的模型與目前最先進的方法相比,有著卓越性能。 Sequential recommendation (SR) takes successive historical behaviors into the next-item prediction to capture dynamic user interests. Recent methods for SR incorporate contrastive learning with self-supervised signals to alleviate the data sparsity problem. While they typically utilize data augmentation to construct contrast signals, the semantic consistency between contrastive objectives has been challenged due to the random perturbation on the original intent. Thus, existing works tend to tackle this problem using a local supervision label such as target item, which is, however, not enough to reflect global sequence behaviors. To further address the above issue, we propose a novel contrastive learning framework, named \textbf{Supe}rvised Pattern Recognition for Sequential \textbf{Rec}ommendation (\textbf{SupeRec}). Different from previous studies, we aim to maintain sequence-level semantic consistency between contrastive objectives by exploring supervised sequence behaviors from raw data. Moreover, a novel contrastive regularization is presented in our work to improve the discrimination ability on false and true negative samples for contrastive tasks. Extensive experiments conducted on five public datasets demonstrate the proposed SupeRec achieve superior performance compared to existing state-of-the-art baselines. |
| URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/88401 |
| DOI: | 10.6342/NTU202302004 |
| Fulltext Rights: | 未授權 |
| Appears in Collections: | 資訊工程學系 |
Files in This Item:
| File | Size | Format | |
|---|---|---|---|
| ntu-111-2.pdf Restricted Access | 3.63 MB | Adobe PDF |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.
