Roundtrip: A Multi-Domain Benchmark for Density Estimation

Roundtrip is a collection of benchmark datasets compiled for evaluating deep generative neural density estimators. It aggregates multiple established datasets from the UCI repository and other sources, covering domains such as activity recognition, protein structure, finance, audio, and particle physics. The collection also includes standard computer vision datasets like MNIST and CIFAR-10, as well as datasets for outlier detection tasks.

Use Cases

• [Density estimation] based on [multi-sensor activity recognition data (AReM) and financial client data (BANK)]
• [Anomaly detection] based on [outlier detection datasets (Shuttle, Mammography, ForestCover) from the ODDS repository]
• [Model evaluation] based on [standard image classification benchmarks (MNIST, CIFAR-10)]
• [Audio feature analysis] based on [song audio data (YPMSD) with 90 features]
• [Regression analysis] based on [protein tertiary structure data (CASP) and high-energy particle data (HEPMASS)]

Strengths

• [Contains large-scale data] with one audio dataset (YPMSD) comprising over 515,000 rows and 90 columns.
• [Provides domain diversity] by aggregating datasets from finance, bioscience, audio, physics, and computer vision into a single benchmark suite.
• [Sourced from established repositories] including the UCI Machine Learning Repository and ODDS, indicating prior use and validation.

Limitations

• [Incomplete metadata] as column names, row counts for some datasets, and last update dates are not provided across all platform entries.
• [Potential format inconsistency] because the original data files are provided in multiple formats (.npy, .csv, .tar.gz).
• [Limited provenance details] for individual datasets beyond their source repository names; specific collection methodologies are not described.

Provenance

Source

Aggregated from the UCI Machine Learning Repository, ODDS repository, and standard vision datasets (MNIST, CIFAR-10).

Collection Method

Datasets were compiled by the author(s) of the 'Roundtrip' research paper for benchmarking purposes.