Delve Datasets for Classification and Regression

A collection of standardized datasets for developing, evaluating, and comparing machine learning methods, maintained by the University of Toronto’s DELVE project.

Overview

• Type: Dataset collection
• Domain: Machine learning (classification and regression)
• Maintained by: University of Toronto, DELVE project
• Primary use: Benchmarking, assessment, and development of learning algorithms
• Access: Downloadable as gzipped-tar files
• Recommended tooling: Delve software environment (for maximum benefit)

Features

Dataset Organization

• Datasets grouped into categories by recommended use:
  • Assessment datasets – for reporting final results; methods should be run once per task without tuning on test data.
  • Development datasets – (mentioned conceptually; used for method development and tuning, details not in excerpt).
  • Historical datasets – (mentioned conceptually as a category; details not in excerpt).
• Within each category, datasets are further labeled as:
  • Regression – continuous target/prototask.
  • Classification – discrete target/prototask.

Access & Format

• Each dataset (or family of datasets) has:
  • A brief overview page.
  • Often detailed documentation (per-dataset docs pages).
• Datasets available as gzipped-tar archives via FTP.
• Installation instructions for downloaded datasets are provided on the site (installation section referenced but not expanded in the excerpt).
• A summary table of all datasets is available for quick reference.

Tooling Integration

• Designed to work with the Delve software environment, which provides:
  • Structured access to datasets.
  • Additional utilities for evaluation (details in a separate "utils" section, not included here).

Dataset Types and Examples

Assessment Regression Datasets

Intended for reporting performance; do not tune on test data.

1. abalone

   • Task: Predict the age of abalone from physical measurements.
   • Source: UCI Machine Learning Repository.
   • Download: abalone.tar.gz (gzipped-tar archive via FTP).

2. bank (bank-family)

   • Type: Family of synthetically generated datasets.
   • Task domain: Simulation of how bank customers choose their banks.
   • Prototask: Predict the fraction of bank customers who leave the bank because of full queues.
   • Download: bank-family (tar archive via FTP).

3. census-house

   • Task: Predict median house prices from 1990 US census data.
   • Download: census-house.tar.gz (gzipped-tar archive via FTP).

4. comp-activ

   • Task: Predict computer system activity from system performance measures.
   • Download: comp-activ.tar.gz (gzipped-tar archive via FTP).

5. pumadyn family of datasets

   • Type: Family of synthetically generated datasets.
   • Task domain: Dynamics of a Unimation Puma 560 robot arm.
   • Description: Generated from a realistic simulation of the robot arm’s dynamics.
   • Download: pumadyn-family (tar archive via FTP).

Assessment Classification Datasets

1. adult

   • Task: Predict whether an individual's annual income exceeds $50,000 based on census data.
   • Source: UCI Machine Learning Repository.
   • Download: adult.tar.gz (gzipped-tar archive via FTP).

2. splice

   • Task: Classification on splice junction data (full description truncated in provided content, but dataset is listed as an assessment classification dataset).
   • Download: splice.tar.gz (gzipped-tar archive via FTP; link partially shown in excerpt).

Documentation & Notes

• An important note for users with version 1.0 of the Delve software is provided on a separate page.
• Each dataset/family has its own desc.html page with additional details (schema, tasks, etc.).

Category

• Directory category: datasets
• Tags: datasets, machine-learning, benchmark

Pricing

• No pricing information is mentioned in the provided content; datasets appear to be freely downloadable from the University of Toronto’s DELVE project site.