AntM2C: A Large-Scale Multi‑Scenario Multi‑Modal CTR Prediction Dataset from Alipay

Click‑through‑rate (CTR) prediction is critical for recommendation systems, influencing user experience and platform revenue. Existing public CTR datasets suffer from limited scenarios, modality, and scale, prompting Ant Group to release the AntM2C dataset, derived from real Alipay industrial data.

AntM2C contains 1 billion CTR samples across five typical Alipay business scenarios (advertising, coupons, mini‑programs, content, and video). Each sample includes ID features and multi‑modal features such as text and images, with over 200 feature columns, making it the largest publicly available CTR dataset.

The first open‑source phase provides 10 million samples (29 ID features, 2 text features) with plans to release the full billion‑sample corpus. The dataset has been de‑identified and encrypted to protect privacy and is intended solely for academic research.

Benchmark tasks include multi‑scenario CTR modeling, cold‑start CTR modeling (few‑shot and zero‑shot), and multi‑modal CTR modeling. Standard AUC is used as the evaluation metric. Results show that multi‑task models outperform simple mixed‑data DNNs, and that incorporating text modalities improves performance in sparse scenarios, highlighting the dataset’s ability to differentiate model capabilities.

Statistical analysis reveals long‑tail distributions for both users and items, with overlapping users across scenarios, reflecting realistic industrial conditions. The dataset also provides auxiliary features such as timestamps and scenario identifiers to facilitate flexible train/validation/test splits.

AntM2C aims to fill the gap in multi‑scenario, multi‑modal CTR research and invites the community to contribute to its ongoing expansion and benchmark development.