Yandex Unveils World's Largest Dataset for Recommender Systems

Yandex has introduced Yambda, the world's largest open dataset for recommender systems, announced in a press release. This dataset includes nearly 5 billion anonymized user interactions from the Yandex Music streaming service, collected over a span of 10 months.

The Yambda dataset is designed to advance research and development in recommender systems by providing a comprehensive benchmark for testing new algorithms. It includes anonymized audio embeddings, interaction flags, and precise timestamps, allowing for detailed real-world behavioral analysis. The dataset is available in three sizes—5 billion, 500 million, and 50 million events—to cater to various research and development needs.

Yambda employs a Global Temporal Split (GTS) evaluation method, which preserves event sequences and provides a realistic testing environment for models. This approach is complemented by baseline algorithms such as MostPop, DecayPop, and ItemKNN, offering reference points for new developments in the field.

The dataset is accessible on Hugging Face, facilitating its use by researchers and startups to innovate and build advanced recommender systems tailored to diverse business needs.