Back to Gap Map Infrastructure / Classic ML & computer vision

PyOD

open source / Overall score: 4.4(strong)

Outlier and anomaly detection library with more than sixty detectors behind one fit/predict API, covering tabular data, time series, graphs and, through foundation-model embeddings, text, images and audio. Version 3 adds ADEngine, which profiles a dataset, picks and runs detectors and assesses the results, and an agent skill and MCP server that drive it from natural-language requests. It is led by Yue Zhao with an advisory committee.

Openness

5 high confidence

5.0

license: BSD-2-Clause(OSI)
source: public(github.com/yzhao062/pyod)
core features withheld: no — academic-led project

BSD-2-Clause and built from the public repository. The project is run by its academic author with an advisory committee and sells nothing; the optional OpenAI encoder calls a paid third-party API but the library itself is whole.

https://pyod.dev/ recorded 2026-09-27
pyod.dev Governance: "Advisory committee: Chris Kuo (Columbia University), Guansong Pang (Singapore Management University), Alicia Guo (IBM Consulting)" and "Every detector and fix lands through pull-request review by the core maintainers"; no pricing or paid tier.
https://raw.githubusercontent.com/yzhao062/pyod/master/LICENSE recorded 2026-09-27
LICENSE is the "BSD 2-Clause License", "Copyright (c) 2018, Yue Zhao".
https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/yzhao062/pyod recorded 2026-09-27
Repository record for yzhao062/pyod, not archived and not a fork, license bsd-2-clause, last pushed 2026-09-17.
https://ungh.cc/repos/yzhao062/pyod/files/master recorded 2026-09-27
Repository tree of 440 paths with no ee, enterprise, commercial or pro directory.

Adoption

4 high confidence

4.0

Measured on monthly PyPI downloads of the pyod package.

https://pypistats.org/api/packages/pyod/recent recorded 2026-09-27
last_month 2,751,853 downloads of pyod

Capability

5 medium confidence

5.0

Beyond its catalog of detectors, PyOD 3 will choose, run and compare detectors on the user's own data without the user picking one, which is the automation AutoGluon offers for prediction. That orchestration layer is what places it here; its detectors alone would sit with the toolkits.

https://raw.githubusercontent.com/yzhao062/pyod/master/README.rst recorded 2026-09-27
README: "PyOD 3: Agentic Anomaly Detection At Scale"; "61 detectors across tabular, time series, graph, text, image, and audio data, one API"; usage layer "ADEngine ... You want PyOD to choose, compare, and assess automatically"; MCP tools "profile_data, plan_detection, build_detector ... run_detection, analyze_results".

Verified 2026-09-27