UCSB_sklearn_workshop

A repository of code and data for the USCB sklearn workshop on February 7th & 8th, 2019

General outline

Day 1

Kaai can arrive early (8:00 AM) to informally talk about using pandas and numpy for reading and processing data if there is interest

Morning 10:00 AM - ~12:00 PM

• Lecture - Basics of machine learning (1-2 hrs)

  • Training and test set
  • Model types, & model complexity
  • Basic data types
  • Data imbalance
  • Performance metrics
  • Advanced data types (NLP, computer-vision, time-series, etc.)

• Code along - First pass: band gap regression (1 hr)

  • download data
  • perform a train-test split
  • convert formulae to features
  • perform linear regression (linear learner)
  • perform random-forest regression (non-linear)
  • generate error metrics & figures

Afternoon 1:00 PM - 5:00 PM

• Code along - Second pass: band gap regression (1.5 hrs)

  • remove duplicates and make sure data looks okay
  • perform a train-test split
  • convert formula to features
  • discuss model parameter selection
  • implement cross-validation
  • implement grid-search
  • generate error metrics & figures

• Individual work - Aflow regression: predicting bluk modulus (until the end of day)

  • Implement code (Taylor/Kaai will be avaliable for questions)
    • Fill out individual code sections. Answers are revealed before moving on (4 or 5 parts).

Homework: Think about data that might be interesting to learn on. We will talk about your ideas in the morning.

Day 2

• Code along (do you want this many coding examples?) - A quick classification problem: predicting crystal structure (1 hr)

  • Augment same code structure for classification
  • generate error metrics & review recal vs precision.

• Individual work - metal/non-metal band gap classification.

  • Work through full ML work flow while
  • Coaching - Discuss research ideas and data with Taylor/Kaai

Bonus material (things we can talk about if there is extra time):

* using matplotlib to make publication quality figures