Scholarlink

<p>Short-video recommendation requires models that capture both evolving user preferences and candidate-video engagement signals. This study proposes V3A-StatFormer, a video statistics-aware sequential recommendation model that integrates Transformer-based user behavior encoding, candidate-video ID matching, and a lightweight embedding module for selected video-level statistical features. The fused representation is used for joint prediction over Video, Click, and LongView objectives. Experiments are conducted on the KuaiRand dataset, which contains 1,306,360 interaction samples and 52 original video statistical fields. Under a unified evaluation protocol, V3A-StatFormer achieves a Test Avg_AUC of 0.8246, outperforming ITEM-CF by 0.0028 and SASRec by 0.0051. It also obtains Recall@10, NDCG@10, and MRR@10 scores of 0.672, 0.452, and 0.384, respectively. Five-seed experiments and paired t-tests further confirm the stability and statistical significance of the improvements. Ablation and feature analyses show that video statistical features provide stable but moderate gains, with engagement-depth and popularity-related features contributing most. These results indicate that compact video statistical features can serve as practical complementary signals for short-video sequential recommendation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A video statistics-aware sequential recommendation model with multi-behavior feedback for short-video recommendation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月6日

查看全文
  • Xinyi Zhou,
  • Huanhuan Xu,
  • Maowei Chen

摘要

Short-video recommendation requires models that capture both evolving user preferences and candidate-video engagement signals. This study proposes V3A-StatFormer, a video statistics-aware sequential recommendation model that integrates Transformer-based user behavior encoding, candidate-video ID matching, and a lightweight embedding module for selected video-level statistical features. The fused representation is used for joint prediction over Video, Click, and LongView objectives. Experiments are conducted on the KuaiRand dataset, which contains 1,306,360 interaction samples and 52 original video statistical fields. Under a unified evaluation protocol, V3A-StatFormer achieves a Test Avg_AUC of 0.8246, outperforming ITEM-CF by 0.0028 and SASRec by 0.0051. It also obtains Recall@10, NDCG@10, and MRR@10 scores of 0.672, 0.452, and 0.384, respectively. Five-seed experiments and paired t-tests further confirm the stability and statistical significance of the improvements. Ablation and feature analyses show that video statistical features provide stable but moderate gains, with engagement-depth and popularity-related features contributing most. These results indicate that compact video statistical features can serve as practical complementary signals for short-video sequential recommendation.