Dataset Viewer
Duplicate
The dataset viewer is not available for this split.
Job manager crashed while running this job (missing heartbeats).
Error code:   JobManagerCrashedError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

AdsTrace: A Multimodal Dataset for Second-by-Second CTR and Advertising Effectiveness Prediction in Short-video Ads

[Paper Link (Coming Soon)] | Project Repository

English | 中文说明


🇬🇧 English Description

💎 AdsTrace Dataset

🛠️ Dataset Tool and TAMAN Framework

Github: https://github.com/XiuzeZhou/AdsTrace/

🌟 Highlights

AdsTrace is a novel, large-scale multimodal benchmark dataset derived from real-world e-commerce short-video advertising. It bridges the gap between fine-grained multimodal content (visual, acoustic, textual) and actual business outcomes.

  • High-Resolution Labels: Provides second-level instantaneous Click-Through Rate (iCTR) trajectories, moving beyond static "black-box" video analysis.
  • Verified Business Metrics: Every video is paired with validated financial outcomes, including Pay ROI and Conversion Rate (CVR).
  • Multimodal Alignment: Synchronized visual frames (1 FPS), extracted acoustic data, and semantic transcripts (Whisper).
  • Statistical Robustness: All 3,000 videos are filtered by a strict exposure threshold ($\tau \ge 3,000$) to ensure reliable performance signals.

📊 Statistics at a Glance

  • Total Samples: 2,833 unique professional advertisements.
  • Domain: E-commerce Ads (Primarily Beauty & Personal Care, e.g., Anti-aging sets, Pro-Xylane series).
  • Duration: Ranges from 6s to 46s (Mean: 21s).
  • Data Distribution:
    • ROI: Near-normal distribution (peak at 1.0–2.0).
    • CVR: Long-tail distribution (peak at 0.28%–0.55%).

📂 Data Structure & Format

To ensure privacy and copyright compliance, we provide high-level features and processed frames:

  • audios_16k/: Audio tracks extracted and resampled to 16kHz mono-channel WAV format (16-bit PCM).
  • frames/: 540x960 JPEG images sampled from raw videos at 1 FPS.
  • ictr: Second-by-second iCTR values of videos.
  • transcripts/: Whisper-transcribed Chinese text.
  • products_cn.json: Chinese names of products.
  • products_en.json: English names of products.
  • tags_cn.csv: Metadata, CVR, and ROI.
  • split.json: Standard Train/Val/Test split indices.

The directory structure is as follows:

├── datasets/AdsTrace/
|   ├── audios_16k/
│   ├── frames/
│   ├── ictr/
│   ├── transcripts/
│   ├── products_cn.json
│   ├── products_en.json
│   ├── tags_cn.csv
│   └── split.json

⚖️ License

This dataset is licensed under CC BY-NC 4.0. Non-commercial research use only.


🇨🇳 中文说明

💎 AdsTrace 数据集

🛠️ 数据集工具与 TAMAN 框架

Github 仓库: https://github.com/XiuzeZhou/AdsTrace/

🌟 核心特点

AdsTrace 是一个源自真实电商短视频广告场景的新型大规模多模态基准数据集。它填补了细粒度多模态内容(视觉、音频、文本)与实际业务结果之间的鸿沟。

  • 细粒度时序标签:提供秒级瞬时点击率 (iCTR) 演化轨迹,超越了传统的静态“黑盒”视频分析。
  • 真实业务指标:每个视频均配有经过验证的真实财务结果,包括 成交 ROI (Pay ROI) 和 **转化率 (CVR)**。
  • 多模态对齐:包含同步的视觉帧(1 FPS)、抽取的音频数据以及由 Whisper 生成的语义转录文本。
  • 统计鲁棒性:所有 3,000 个视频均通过严格的曝光阈值筛选($\tau \ge 3,000$),以确保效果信号的可靠性。

📊 数据概览

  • 总样本量:2,833 条独特的专业广告视频。

  • 覆盖领域:电子商务广告(主要集中在美妆护肤,如抗衰套装、玻色因系列等)。

  • 视频时长:分布在 6s 至 46s 之间(平均时长 21s)。

  • 数据分布:

    • ROI:近似正态分布(峰值处于 1.0–2.0 之间)。

    • CVR:长尾分布(峰值处于 0.28%–0.55% 之间)。

📂 数据结构与格式

为了遵守版权和隐私政策,我们提供处理后的多模态特征: 为符合隐私保护和版权合规要求,我们提供高层特征及处理后的视频帧:

  • audios_16k/:提取并重采样至 16kHz 单声道 WAV 格式的音频轨道(16-bit PCM)。
  • frames/:从原始视频中按 1 FPS 采样得到的 540x960 JPEG 图像。
  • ictr/:视频每秒对应的瞬时点击率数值。
  • transcripts/:使用 Whisper 转录的中文文本。
  • products_cn.json:产品中文名称。
  • products_en.json:产品英文名称。
  • tags_cn.csv:元数据、CVR 及 ROI 指标。
  • split.json:标准的训练/验证/测试集划分索引。

目录结构如下:

├── datasets/AdsTrace/
|   ├── audios_16k/
│   ├── frames/
│   ├── ictr/
│   ├── transcripts/
│   ├── products_cn.json
│   ├── products_en.json
│   ├── tags_cn.csv
│   └── split.json

⚖️ 许可协议

本数据集采用 CC BY-NC 4.0 协议。仅限非商业性研究使用。


Downloads last month
50