Hello Austin Data Meetup!

What this talk is about

• Machine learning evaluation is a pain, and in my experience we're all terrible at it
• Here's what we should do instead
• Say hello to LeVar, a database designed to help us do what we should do

Some more about LeVar

• Demo!
• A bit about how it's implemented, if you're interested
• Open issues and next steps

Machine learning evaluation is a pain and we're all terrible at it

Evaluation is important

• Evaluate early, evaluate often
• Keep a lab notebook
• Evaluate every change
• Do error analysis
• Change your model based on evidence

But what happens

• No standard source, storage or format for eval
• End up rewriting our evaluation scripts for each project
• Often couple our evaluation code with our core ML code
• Error analysis is a pain
• Hard to track results on the same evaluation over time
• Harder to to comparative error analysis over time

Our solution

We can do better

• Keep evaluation datasets in a centralized location
• Evaluations are largely immutable
• Not tied to any one ML framework (e.g. scikit)
• ...or big data framework (e.g. RDDs)
• Totally agnostic WRT to method
• Human-readable datasets
• Use simple and common format for data exchange
• Open schema for data points, i.e. arbitrary features

We can do better

• Command-line tool for data import & power users
• Several standard problem schema supported
  • Classification
  • Regression
  • Geo-prediction
  • Structure prediction
  • Machine translation
• Several standard useful evaluation criteria supported
  • Accuracy
  • Precision/recall/F-score
  • ROC
  • RMSE...

We can do better

• Web UI showing high-level experiment results suitable for bosses or clients
• Web UI & CLI search for error analysis
• User can comment on anything:
  • Dataset
  • Experiment
  • Item in dataset
  • Individual prediction in experiment
  • Another comment
• User can label anything

We can do better

• Sensible information architecture
  • Organize datasets into groups or organizations
• Provide sensible baseline evaluations out-of-the-box
  • Most-common class
  • Mean value
  • (Weighted) random

We can do better

• REST API to use with your favorite framework
• Straightforward client libraries
• Export and import to other formats (yo RDD)

Introducing LeVar

It does several of those things!

But don't freak it's just v0.1

Here's what's done now

• CLI (in Scala)
• Import/export TSV
• Organizations data model
• API for everything, using basic auth
• Client code for Scala
• Classification & regression datasets
• Many useful evaluations

LeVar data model

• Datasets
• Items (datum)
• Experiments
• Predictions

Everything has a creation date; everything has a UUID

The data model should support most of those other features, so I can already do e.g. error analysis by dropping into psql and writing a query

On Github

https://github.com/peoplepattern/LeVar

Demo?

Future directions

Let's just jump over to the issues page

https://github.com/peoplepattern/LeVar/issues

THANKS!