From Lab to Factory

Lessons learned developing Data Products

Keynote PyCon Colombia

Feb 2017

peadarcoyle@googlemail.com

All opinions my own

Who am I?

Data Scientist at Elevate (Recruitment Startup) and Author

@springcoil

About 5 years in Machine Learning/Data Science

OSS contributor (mostly PyMC3)

Built Ad-hoc analysis, Data Products and Data Pipelines at

Why talk about this?

Data Moats

Roadmap

What is Data Science?

Type of Data Scientists

Deploying Data Science (tech/culture)
Success stories and recommendations
Bigger idea (Change)

What IS a Data Scientist?

There's a Data Science Spectrum

HT: Sean J. Taylor (Facebook Research Scientist)

Data Science to Value

Data doesn't do anything!
People need to make decisions (strategic/ tactical)
Or the product needs to alter someones decision making (Spotify Discover Weekly, Amazon Recommendations, etc)

(Taken from Josh Wills talk - Lab to Factory 2014)
or https://hbr.org/2013/04/two-departments-for-data-succe

Expectations and reality

Data is the new oil...

But. Data isn't a product. It needs to be turned into one!

Data Science is hard

Data and technical problems

Skill set (emotional and tech)

Machine Learning Ops

Data Science projects are risky!

Many stakeholders think that data science is just an engineering problem, but research is high risk and high reward

http://www.martingoodson.com/ten-ways-your-data-project-is-going-to-fail/

De-risking the project - how? Send me examples :)

https://github.com/ianozsvald/data_science_delivered

Success with Data Science

1. People

2. Ideas

3. Things

Source: USAF

R and D needs cultural support

Your culture needs to avoid the HiPPO (highest paid persons opinion) to be data-informed or data-driven.

Good cultures

Data democratization
Clear objectives and metrics against objectives
Financial metrics don't always win - sometimes brand wins
See http://bit.ly/cultureanddata by Martin Goodson

Some data scientists do experiments and build prototypes

Production

(HT: The Yhat people - www.yhathq.com)

Deploying Data Products is hard

Monitoring and Alerting
Deployment
Real world data is tricky (training/ testing)
Interpretability (explaining fraud detection/ad tech)
Feature engineering
Models decay over time

What projects work?

Explain existing data (visualization!)
Automate repetitive/ slow processes
Augment data to make new data (Search engines, ML models)
Predict the future (do something more accurately than gut feel )
Simulate using statistics :) (Sports analytics models)

"You need data first" - Peadar Coyle

Copying and pasting PDF/PNG data
Getting data in some areas is hard!!
Some tools for web data extraction
Messy APIs without documentation :(

Visualise ALL THE THINGS!!

(Relay foods dataset - HT Greg Reda)
Consumer behaviour at a Fast Food Restaurant per year in the USA

Augmenting data and using API's

Machine Learning

Production data/ Real world

HT: Andrew Ng

Everyone ETLS

Only 1% of your time will be spent modelling
Data pipelines and your infrastructure matters - Eoin Brazil Talk
Tools such as Spark, Luigi, Airflow or Dask are important
Monitoring of pipelines. Scalability. Machine provision

Success Story (1)

Ad Tech project (Media company) -
original process took 10 hours per week of analyst time.
Broke down process and produced simple predictive model
Data type challenges (unicode in Python 2.7) and business expectations
Now takes less than 30 minutes each week.

Success Story (2)

Elevate is a Total Talent Management platform (tools for streamlining recruitment)
Several models in production - microservices (recommenders, job type prediction)
Data pipeline (Luigi) necessary for producing features
Lots of challenges about Docker, time, updated data.
Works: MyPy (Python 3.5), code review, testing - http://hypothesis.works/
Next: Moving to PySpark for ETL, more pair programming

Failure

1. Lack of support from tech teams

2. Lack of clearly defined customer

3. No culture of DevOps

4. Management had no clear vision

Lessons learned from Lab to Factory

1. Monitoring - data products need evaluation in production

2. Lack of a shared language between software engineers and data scientists. Pair programming/ Code review are some solutions.

3. To help data scientists and analysts succeed your business needs to be prepared to invest in tooling. (Pipelines, DevOps, Reproducible)

4. Often you're working with other teams who use different languages - so micro services can be a good idea (example C#.net app and Python microservice)

How to deploy a model?

Palladium (Otto Group)

Azure

Flask Microservice (DIY) and Docker

Dev Ops

We need each other!

Use small data where possible!!

Small problems with clean data are more important - (Ian Ozsvald)
Amazon machine with many Xeons and 244GB of RAM is less than 3 euros per hour. - (Ian Ozsvald)
Blaze, Xray, Dask, Ibis, etc etc - PyData Bikeshed
"The mean size of a cluster will remain 1" - Matt Rocklin

Focus on the real problem

Closing remarks

Dirty data stops projects
Change happens, we need to help businesses adapt or be agile
There are some good projects like Icy, Luigi, etc for transforming data and improving data extraction
http://martin.zinkevich.org/rules_of_ml/rules_of_ml.pdf - Google paper on rule for building ML systems