Hey, Listen! This presentation works THE BEST in Webkit/Blink browsersonly. Sorry about that. Pull requests welcome?

AutoQC

A new python framework for large scale data quality control

the state of open data

- Gov'ts, granting agencies and journals demanding data be released publicly

- NOAA already offering extensive open data services

- Evidence mounting for tremendous scientific gains from sharing data

- But how to make sure open data actually pays off?

IQuOD can lead the scientific community

- Automatic data QC will be necessary for a healthy open data community

- Have the opportunity to build infrastructure to help correctly interpret open data

- Can best succeed at this project via an open source collaboration.

Introducing AutoQC

Challenge:

- Pass or fail ocean profiles based on sanity checks

- ...simple enough?

Problems:

- Prior Art

- Scale & Speed

- Decision Making

Prior Art

- you've already written a ton of sanity checks

- but, they're in C, C++, FORTRAN...

- how do we make it easy to combine past work under one roof?

Modularity First

- AutoQC is 100% plug & play

- AutoQC provides standardized WOD unpacking

- Includes testing framework to ensure code quality

- Allow me to demonstrate...

Introducing wodpy

- supports python 2.7/3.x

- supports numpy & pandas data formats

- highly unit tested for reliability

- open source & available on PyPI

pip install wodpy

Scale & Speed

- quota alone has > 150k profiles

- takes ~24h to run AutoQC without parallelization on this laptop

- how can we run more tests on more data for cheap?

Parallelism First

- AutoQC easy to parallelize

- Factored to record qc results before postprocessing

- Demonstrated on AWS c3.8xLarge cluster:

- 150k profiles analyzed in 30 minutes @ $1.68 USD

Decision Making

- No qc test catches all bad profiles

- All qc tests produce false positives

- Could some qc tests be redundant?

- What is the minimal & optimal combination of qc tests to classify a profile?

AutoQC Performance Testing

- Examine ~150k profiles from NOAA's quota dataset

- AutoQC currently implements the following QC tests:

  • Argo_global_range_check
  • Argo_gradient_test
  • Argo_impossible_date_test
  • Argo_impossible_location_test
  • Argo_pressure_increasing_test
  • Argo_regional_range_test
  • Argo_spike_test
  • EN_background_check
  • EN_constant_value_check
  • EN_range_check
  • EN_spike_and_step_check
  • EN_stability_check
  • WOD_gradient_check
  • WOD_range_check

'Or' Everything Together

- If we flag a profile when any of the underlying tests do so, we find:

  • 54.8% of bad profiles flagged
  • 90.4% of good profiles passed
  • 45.2% false negative rate
  • 9.6% false positive rate
To-Do #1: need more QC tests to eliminate zero-flag spike

EN-Background

Compared to a 54.8% flag rate of or'ing everything together, EN-Background produced

  • 49.0% of bad profiles flagged
  • 93.4% of good profiles passed
  • 51.0% false negative rate
  • 6.6% false positive rate

Is quota a biased sample?

To-Do #2: need to better understand origin of flags in quota, to determine if the underlying dataset is biasing these results.

Machine Learning

- Once we hammer out to-do's 1 & 2, we'll need to make final flagging decisions.

- OR'ing everything together is probably a good start...

- The optimal decision strategy will likely be a machine learning algorithm.

Support Vector Machines

- AutoQC is essentially an attempt to learn a boolean function

- Ken Sadohara has presented promising work on using SVMs for this purpose

- scikit-learn provides powerful and easy-to-use tools for doing machine learning in python

Get Involved

- Help implement more QC tests; it's easy!

- Help construct a large representative sample of profiles to gauge performance on.

AutoQC

@billdoesphysics

http://billmills.github.io/IQUOD-Hamburg/slides/

https://github.com/IQuOD/AutoQC