[4O1-IS-2a] Machine learning

9:15 AM - 9:30 AM JST(12:15 AM - 12:30 AM UTC)

[4O1-IS-2a-02] A video-oriented recommendation system

Ying Yu Chen¹, 〇WEI CHI WANG², Lieu Hen Chen², Yasufumi Takama¹ (1. Univ. of Tokyo Metropolitan, 2. National Chi Nan University)

regular

Keywords:

Video-Oriented Recommendation,Graphically Recommendation Interface,Shikakeology,Temporal Alignment,Flip Analysis

In the era of information explosion, AI-based recommendation systems face the dual challenge of mitigating user aversion to advertisements and ensuring precise content alignment. We propose a graphically illustrated recommendation interface incorporating Shikakeology strategies, where recommendations are embedded within a tarot-reading metaphor. In recommendation systems, ranking determines the final recommendation outcome; however, for video-oriented content, static representations fail to capture users’ dynamic intentions. To address this limitation, we introduce a video-oriented recommendation framework that incorporates temporal characteristics into the ranking process. Based on users’ emotional intentions, video candidates are ranked by evaluating alignment across semantic, affective, and temporal dimensions. We further introduce the Temporal Intent Token (TIT) as an interpretable feature modeling pacing and motion. Experimental results show that temporal alignment substantially influences recommendation decisions, yielding an approximately 83% flip rate in top-1 recommendations compared to static baselines, demonstrating the effectiveness of our approach for generative video recommendation scenarios.