Phishing Websites: Subsampled Dataset for Machine Learning

A subsampled version of the PhishingWebsites dataset from OpenML, created with a specific random seed. The subset contains up to 2000 rows, 100 columns, and 10 classes, generated via stratified sampling. The original dataset is in the public domain, authored by Eddie Bergman.

Use Cases

• Training binary or multi-class classifiers to identify phishing websites based on URL and page features.
• Benchmarking feature selection algorithms on a dataset with a controlled number of columns.
• Evaluating model performance on stratified subsets to ensure class distribution is maintained.
• Studying the impact of dataset size (row count) on phishing detection model accuracy.

Strengths

• Subsampling parameters are explicitly documented: seed=2, nrows=2000, ncols=100, nclasses=10.
• Stratified sampling was used (no_stratify=False), which likely preserves the original class distribution in the subset.
• Derived from a known source dataset (PhishingWebsites ID 4534) on OpenML, providing traceability.

Limitations

• The exact column names, data types, and semantic meanings are not provided in the metadata.
• Row count is specified as a target (2000) but the actual final count after stratified sampling is unknown.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

OpenML dataset 'PhishingWebsites' (ID 4534), subsampled by Eddie Bergman.

Collection Method

Programmatic subsampling using a Python script with random seed-based selection of rows, columns, and classes.