On-Point-Rnd/ESdB-Embeddings-for-Sequential-data-Benchmark

ESdB: Embeddings for Sequential Data Benchmark

ESdB provides reproducible splits, evaluation shifts, and downstream targets for benchmarking representations of sequential data.

This repository contains benchmark annotations only. It does not redistribute the original events or input features. Original datasets must be obtained from their respective sources and can be reproduced with the preprocessing code in the ESdB repository.

Structure

Each dataset contains train and test Parquet directories:

<dataset>/train/
<dataset>/test/

The files contain:

• an identifier inherited from the source dataset;
• shifts: positions used to construct local downstream targets;
• global_train: the global-task train/test assignment;
• target__*: source-provided or ESdB-derived downstream targets.

Datasets

30music, age, alpha, electric_devices, ett, favorita, rossman, taobao, twitter, x5-retail, yambda, and zvuk.

License

CC BY 4.0 applies to the benchmark annotations, split definitions, evaluation shifts, and documentation created by the ESdB authors. It does not relicense the original source datasets. Source-provided identifiers and labels remain subject to their respective source terms.