Earticle

초록

영어

Real-time interaction between anchors and viewers is a defining feature of live streaming e-commerce, shaping re-lational engagement and immediate purchase decisions. However, existing sales prediction models largely rely on aggregated behavioral metrics, overlooking the temporal and reciprocal dynamics that drive sales outcomes. Ana-lyzing 7,684 broadcasts, we find that sales performance depends on distinctive temporal trajectories of interaction rather than static engagement levels. To address this, we propose AVITSNet (Anchor–Viewer Interaction-aware Time Series Network), a multimodal deep learning framework that explicitly models the temporal and bidirectional flow of anchor–viewer interactions. Experimental results show that AVITSNet consistently outperforms conventional and hybrid baselines across all metrics. Textual signals (anchor speech and viewer comments) serve as key predictors, while be-havioral and contextual variables provide complementary value. Attention analysis further highlights the predictive importance of early viewer responses.

목차

Abstract
1. Introduction
2. Related Work
2.1 Interactive Engagement in Live Commerce
2.2 Live Streaming E-Commerce Sales Prediction
3. Data and Descriptive Analysis
3.1 Context Data
3.2 Time-Series Data
3.3 Descriptive Patterns of Interaction and Sales
4. Framework
4.1 Sequential and Cross-Stream Modeling
4.2 Fusion and Prediction
5. Experiments
5.1 Overall Model Comparison
5.2 Feature Ablation Study
5.3 Attention-Based Interpretation
6. Discussion and Conclusion
Acknowledgments
References

저자

• 정도현 [ KAIST 데이터사이언스대학원 ]
• 진승욱 [ KAIST 경영공학부 ]
• 김지영 [ 성균관대학교 경영대학 ]
• 강금석 [ KAIST 경영공학부 ]

참고문헌