What Types of Machine Learning Tasks and Domains Are Covered by UCI Machine Learning Repository Datasets?

The UCI Machine Learning Repository covers a broad range of machine learning tasks and application domains, with classification and regression being the most visible task types among its popular datasets. According to the repository's homepage, it currently maintains 689 datasets, spanning domains such as healthcare, finance, marketing, education, and audio signal processing, with dataset sizes ranging from 150 instances to over 3 million. This makes it suitable for academic research, algorithm benchmarking, and teaching.

Task Types Represented

The repository's popular and new dataset listings explicitly label datasets with task types. Based on the homepage evidence, the following task types appear:

Classification is the most frequently listed task type among the highlighted datasets. Regression appears both as a standalone task (Amazon/Google Reviews) and combined with classification (Wine Quality, Student Performance). The homepage excerpt does not explicitly list clustering or other unsupervised task types among the featured datasets, so those should not be assumed from this evidence alone.

Application Domains Covered

The featured datasets illustrate several domains:

• Healthcare and medicine: Heart Disease (4 databases: Cleveland, Hungary, Switzerland, and VA Long Beach), Breast Cancer Wisconsin (Diagnostic), and HLS-CMDS (heart and lung sounds recorded from a clinical manikin using a digital stethoscope).
• Finance and marketing: Bank Marketing, which relates to direct marketing campaigns (phone calls) of a Portuguese banking institution, with the goal of predicting whether a client will subscribe a term deposit.
• Food and beverage quality: Wine Quality, covering red and white vinho verde wine samples from northern Portugal, modeling quality based on physicochemical tests.
• Education: Student Performance, predicting performance in secondary education (high school).
• E-commerce and reviews: Amazon Product and Google Locations Reviews, a preprocessed dataset with time series of hourly review counts across categories.
• Classic benchmarking: Iris, a small classic dataset from Fisher (1936), described as one of the earliest known datasets for evaluating classification methods.

Dataset Scale

Dataset sizes vary widely, which affects what each is suited for:

• Small (hundreds of instances): Iris (150 instances, 4 features), Heart Disease (303 instances, 13 features), Breast Cancer Wisconsin (569 instances, 30 features), Student Performance (649 instances, 33 features).
• Medium (thousands): Wine Quality (4.9K instances, 12 features).
• Large (tens of thousands): Bank Marketing (45.21K instances, 17 features).
• Very large (millions): Amazon Product and Google Locations Reviews (3.06M instances, 2 features).

Small datasets like Iris and Heart Disease are commonly used for teaching and quick algorithm comparisons. Larger datasets such as Bank Marketing and the Amazon/Google Reviews dataset are better suited for testing scalability or working with higher-volume data.

Choosing a Dataset for Your Task

• For classification benchmarking or teaching: Iris, Heart Disease, or Breast Cancer Wisconsin offer small, well-known classification problems.
• For regression or combined classification/regression: Wine Quality and Student Performance are labeled for both task types.
• For imbalanced or real-world marketing problems: Bank Marketing provides a larger classification dataset tied to a concrete business goal.
• For large-scale or time series work: The Amazon Product and Google Locations Reviews dataset offers millions of instances, though with only 2 features.
• For audio or signal-based tasks: HLS-CMDS provides heart and lung sound recordings, including individual and mixed sounds with normal and abnormal examples.

The repository also allows users to browse all datasets and contribute their own, so the coverage extends beyond the featured examples. To confirm the exact task type, domain, instance count, and feature count for any dataset, check its individual page on the repository.