簡易檢索 / 詳目顯示
| 研究生： |
朱茂榛 Chu, Mao-Chen |
|---|---|
| 論文名稱： |
基於動態三方超圖網路之直播推薦系統 Dynamic Tripartite Hypergraph Networks for Live-Streaming Recommendation |
| 指導教授： |
蕭舜文
Hsiao, Shun-Wen 林怡伶 Lin, Yi-Ling |
| 口試委員: |
陳孟彰
Chen, Meng-Chang |
| 學位類別： |
碩士
Master |
| 系所名稱： |
商學院 - 資訊管理學系 Department of Management Information System |
| 論文出版年： | 2026 |
| 畢業學年度： | 114 |
| 語文別： | 英文 |
| 論文頁數： | 84 |
| 中文關鍵詞： | 動態圖學習 、直播電商 、推薦系統 、三方圖 |
| 外文關鍵詞： | Dynamic graph learning, Live-streaming e-commerce, Recommender System, Tripartite Graph |
| 相關次數： | 點閱：22 下載：0 |
| 分享至: |
| 查詢本校圖書館目錄 查詢臺灣博碩士論文知識加值系統 勘誤回報 |
直播電商的一次互動，同時牽涉使用者、直播主與商品三方，且變動快速。然而現有的連續時間動態圖模型建立在成對且同質的假設上，會將這樣的事件拆解為彼此孤立的二元關係，因而喪失其聯合語意。本研究將直播互動重新形式化為時序三方超圖，並提出異質三方轉換器（Heterogeneous Tripartite Transformer, HT-Transformer）：完整保留每次事件為不可分割的超邊，以角色專屬的編碼器建模三方的歷史共現關係，並由使用者條件化的閘門機制，依其行為集中程度重新調整直播主與商品表徵的權重。於 KuaiLive 資料集上，HT-Transformer 的表現優於靜態、異質與連續時間等各類基準模型，在嚴格的前段排序指標上差距尤其明顯。分層分析進一步顯示，此優勢在每一個直播主群組中皆成立，且當直播主自身的歷史資訊越稀少時，優勢越為擴大。本研究並刻畫純結構式歸納推論的能力邊界：不具歷史紀錄的直播主是最困難的冷啟動情境，而在無須重新訓練、亦不依賴內容特徵的前提下，借用「該使用者原本即已關注之既有直播主」的互動脈絡，即可挽回部分效能損失。
A live-streaming e-commerce interaction couples a user, a streamer and an item in a single, fast-moving event. Continuous-time dynamic graph models, however, rest on pairwise and homogeneous assumptions: they decompose such an event into isolated dyads and lose its joint semantics. We reformulate live-streaming interactions as a temporal tripartite hypergraph and propose the Heterogeneous Tripartite Transformer (HT-Transformer), which preserves each event as an indivisible hyperedge, encodes the historical co-occurrence of the three roles with role-specific encoders, and lets a user-conditioned gate reweight the streamer and item representations according to the user’s behavioural concentration. On the KuaiLive dataset, HT-Transformer outperforms static, heterogeneous and continuous-time baselines, most clearly on strict top-rank metrics. A stratified analysis shows that this advantage holds within every streamer group and widens as the streamer’s own history becomes less informative. We further characterize the limits of purely structural inductive inference: a streamer without history is the hardest cold-start case, and part of the loss is recovered—without retraining or content features—by borrowing the interaction context of a warm streamer that the querying user already follows.
誌謝 i
摘要 ii
Abstract iii
Contents iv
List of Figures vii
List of Tables viii
1 Introduction 1
2 Background and Related Work 6
2.1 Live-streaming E-commerce 6
2.2 Graph Structures and Temporal Dynamics 7
2.3 Related Works in Recommendation Systems 8
2.4 Continuous-Time Dynamic Graph Learning 9
2.5 Sparsity and Generalization in Dynamic Graphs 11
2.6 Summary and Research Gap 12
3 Research Design 14
3.1 From Research Gaps to Design Decisions 14
3.2 Key Design Decisions and Rejected Alternatives 14
3.3 Expected Behavior and Scope of the Design 17
4 Model Architecture 19
4.1 Problem Formalization: Temporal Tripartite Hypergraph 21
4.2 Type-aware Node Initialization 23
4.3 3D Neighbor Co-occurrence Encoding 24
4.4 Sequence Patching for Tripartite Interactions 25
4.5 Tripartite Attention and Representation Learning 26
4.6 Bias-Aware Tripartite Merge Layer 28
5 Experiments 30
5.1 Experimental Objectives and Research Questions 30
5.2 Dataset and Statistical Analysis 31
5.3 Experimental Task Setting 38
5.4 Baselines and Compared Variants 39
5.5 Evaluation Metrics 42
5.6 Implementation Details 44
6 Results and Analysis 46
6.1 Main Performance Comparison 46
6.2 Ablation Study on the Bias-Aware Gate 49
6.3 Ablation Study on Heterogeneous Modeling Components 50
6.4 Bias-Aware Effect under Different User Groups 53
6.5 Generality Across Streamer Types 55
6.6 Necessity of Tripartite Modeling (Modality Ablation) 58
6.7 Impact of Temporal Information: A Two-Level Analysis 61
6.8 Temporal Dynamics and Ranking Crossover Analysis 63
6.9 Cold-Start Scenario Analysis 65
6.10 Top-Ranked Architecture Analysis 68
6.11 Linking Cold Streamers to Warm Ones 70
6.12 Phenomenon-Driven Analysis 73
6.13 Hyperparameter Sensitivity Analysis 74
7 Conclusions 76
7.1 Summary of Findings 76
7.2 Research Limitations 78
7.3 Future Work 80
References 81
Cong, W., Zhang, M., Chen, C., and Zhou, D. (2023). Do we really need complicated model architectures for temporal networks? In Proc. Int. Conf. Learn. Represent. (ICLR).
Fan, X., Zhang, L., Guo, X., and Zhao, W. (2024). The impact of live-streaming interactivity on live-streaming sales mode based on game-theoretic analysis. J. Retail. Consum. Serv., 81:103981.
Hamilton, W. L., Ying, R., and Leskovec, J. (2017). Inductive representation learning on large graphs. arXiv preprint arXiv:1706.02216.
He, X., Deng, K., Wang, X., Li, Y., Zhang, Y., and Wang, M. (2020). Lightgcn: Simplifying and powering graph convolution network for recommendation. In Proc. 43rd Int. ACM SIGIR Conf. Res. Develop. Inf. Retr. (SIGIR), pages 639–648.
Hirschman, A. O. (1964). The paternity of an index. Am. Econ. Rev., 54(5):761–762.
Hu, Z., Dong, Y., Wang, K., and Sun, Y. (2020). Heterogeneous graph transformer. In Proc. World Wide Web Conf. (WWW), pages 2704–2710.
Jiang, J., Wei, Y., Feng, Y., Cao, J., and Gao, Y. (2019). Dynamic hypergraph neural networks. In Proc. 28th Int. Joint Conf. Artif. Intell. (IJCAI), pages 2635–2641.
Kang, W.-C. and McAuley, J. (2018). Self-attentive sequential recommendation. In Proc. 18th IEEE Int. Conf. Data Mining (ICDM), pages 197–206.
Kazemi, S. M., Goel, R., Eghbali, S., Ramanan, J., Sahota, J., Thakur, S., Wu, S., and Smyth, C. (2019). Time2vec: Learning a vector representation of time. In Adv. Neural Inf. Process. Syst. (NeurIPS), volume 32.
Kumar, S., Zhang, X., and Leskovec, J. (2019). Predicting dynamic embedding trajectory in temporal interaction networks. In Proc. 25th ACM SIGKDD Int. Conf. Knowl. Discovery Data Mining (KDD), pages 1269–1278.
Lu, Z., Xia, H., Heo, S., and Wigdor, D. (2018). You watch, you give, and you engage: A study of live streaming practices in china. In Proc. CHI Conf. Hum. Factors Comput. Syst. (CHI), pages 1–13.
Mei, H. and Eisner, J. M. (2017). The neural hawkes process: A neurally self-modulating multivariate point process. In Adv. Neural Inf. Process. Syst. (NeurIPS), pages 6757–6767.
Peintner, A., Mohammadi, A. R., Müller, M., and Zangerle, E. (2025). Hypergraph-based temporal modelling of repeated intent for sequential recommendation. In Proc. ACM Web Conf. (WWW), pages 3714–3723.
Qu, C., Dai, S., Guo, K., Zhao, L., Niu, Y., Zhang, X., and Xu, J. (2025). Kuailive: A real-time interactive dataset for live streaming recommendation. arXiv preprint arXiv:2508.05633.
Rossi, E., Chamberlain, B., Frasca, F., Eynard, D., Monti, F., and Bronstein, M. (2020). Temporal graph networks for deep learning on dynamic graphs. arXiv preprint arXiv:2006.10637.
Schein, A. I., Popescul, A., Ungar, L. H., and Pennock, D. M. (2002). Methods and metrics for cold-start recommendations. In Proc. 25th Annu. Int. ACM SIGIR Conf. Res. Develop. Inf. Retr. (SIGIR), pages 253–260.
Sun, Y., Shao, X., Li, X., Guo, Y., and Nie, K. (2019). How live streaming influences purchase intentions in social commerce: An it affordance perspective. Electron. Commer. Res. Appl., 37:100886.
Trivedi, R., Farajtabar, M., Biswal, P., and Zha, H. (2019). Dyrep: Learning representations over dynamic graphs. In Proc. Int. Conf. Learn. Represent. (ICLR).
Wang, J., Ding, X., Hong, L., Liu, H., and Wang, J. (2020). Next-item recommendation with sequential hypergraphs. In Proc. 43rd Int. ACM SIGIR Conf. Res. Develop. Inf. Retr. (SIGIR), pages 1101–1110.
Wang, L., Chang, X., Li, S., Chu, Y., Li, H., Zhang, W., He, X., Song, L., Zhou, J., and Yang, H. (2021a). Tcl: Transformer-based dynamic graph modelling via contrastive learning. arXiv preprint arXiv:2105.07944.
Wang, X., Ji, H., Shi, C., Wang, B., Ye, Y., Cui, P., and Yu, P. S. (2019). Heterogeneous graph attention network. In Proc. World Wide Web Conf. (WWW), pages 2022–2032.
Wang, Y., Chang, Y.-Y., Liu, Y., Leskovec, J., and Li, P. (2021b). Inductive representation learning in temporal networks via causal anonymous walks. In Proc. Int. Conf. Learn. Represent. (ICLR).
Wongkitrungrueng, A. and Assarut, N. (2020). The role of live streaming in building consumer trust and engagement with social commerce sellers. J. Bus. Res., 117:543–556.
Wu, S., Tang, Y., Zhu, Y., Wang, L., Xie, X., and Tan, T. (2019). Session-based recommendation with graph neural networks. In Proc. 33rd AAAI Conf. Artif. Intell. (AAAI), pages 346–353.
Xu, D., Ruan, C., Korpeoglu, E., Kumar, S., and Achan, K. (2020). Inductive representation learning on temporal graphs. In Proc. Int. Conf. Learn. Represent. (ICLR).
Yu, L., Sun, L., Du, B., and Lv, W. (2023). Towards better dynamic graph learning: New architecture and unified library. In Adv. Neural Inf. Process. Syst. (NeurIPS), volume 36, pages 67686–67700.
Yu, S., Jiang, Z., Chen, D., Feng, S., Li, D., Liu, Q., and Yi, J. (2021). Leveraging tripartite interaction information from live stream e-commerce for improving product recommendation. In Proc. 27th ACM SIGKDD Conf. Knowl. Discovery Data Mining (KDD), pages 2360–2368.
全文公開日期 2031/08/18