Machine learning for particle physics using R – Data production pipeline – Visualising a correlation matrix as a force-directed network

Machine learning for particle physics using R

Andrew John Lowe

Wigner Research Centre for Physics,

Hungarian Academy of Sciences

Introduction: about this talk

• I'm shamelessly recycling slides that I presented at a data science conference in Budapest two weeks ago, so what follows will contain little (if any) physics
• As this is the inaugaral meeting of the LHC machine learning WG, let's start with the absolute basics: tools
• I'll talk about how using R has made it easier for me to ask more complex questions from my data than I would have been able to otherwise
• I'm not advocating that everyone should switch to R
• Hopefully this talk will be complementary to discussions of ROOTR and RMVA

Tools

• Recognise that when you chose a specific analysis tool or framework, you are making a choice that invariably involves some kind of trade-off
  • Not always obvious what this compromise is
  • Do you have the right tool for the job?
• When your only tool is a hammer, everything looks like a nail
  • Generally a good idea to have some familiarity with more than one data analysis tool
• What analysis tools should our students learn?
  • Given that PhD \(\neq\) academic career, what knowledge of tools would provide a good RoI and best equip students with the technical skills they need in a job outside HEP?
• What is available to particle physicists?

What happened outside HEP in the past 20 years?

On C++ and data analysis

• Is C++ a good choice for data analysis?
  • Spend days coding something that runs in minutes or
  • Write something in a couple of hours that will run during your lunch break?
  • Which will get you your answer faster? What strategy will help you define where you should be focusing your efforts and which paths lead to dead-ends?
• Larry Wall, creator of Perl (speaking about differences in the number of lines of code needed to accomplish the same task using different languages):

You can eat a one-pound stake, or a 100 pounds of shoe leather, and you feel a greater sense of accomplishment after the shoe leather, but maybe there's some downsides...

What I want in my tool box:

Sometimes I want to use:

Or sometimes:

What ROOT provides

There are positives and negatives here...

Why did I choose R?

• Chief among those were the need for fast prototyping and high-level abstractions that let me concentrate on what I wanted to achieve, rather than on the mechanics and the highly-granular details of how I might do it
• Incredibly easy to express what I want to achieve
• Exponentially-growing number of add-on packages
• Latest machine learning algorithms are available
• About 2 million R users worldwide\(^*\); technical questions are answered extremely quickly (if not already)
• Not as fast as C++, but my goal is to quickly test new ideas rather than implement a final model
• Beautiful plots
• Fun to work with ☺

* http://www.inside-r.org/what-is-r

Downsides to using R?

• Can be very slow
  • Base R is single-threaded
  • Unlikely to do a full HEP analysis with all the data in R!
  • But fine for looking at small chunks of data
  • Avoid for loops
• Your data has to fit into RAM
  • But not always strictly true
    • Revolution R Enterprise provides the RevoScaleR package and XDF file format for Big Data
    • H2O.ai enables users to import data via a reference to an object in a distributed key-value store
    • Other packages: ff, bigmemory...

Getting ROOT data into R

• RootTreeToR enables users to import ROOT data into R
• Written by Adam Lyon (FNAL), presented at useR! 2007
• cdcvs.fnal.gov/redmine/projects/roottreetor
• Requires ROOT to be installed, but no need to run ROOT

# Open and load ROOT tree:
rt <- openRootChain("TreeName", "FileName")
N <- nEntries(rt) # number of rows of data
# Names of branches:
branches <- RootTreeToR::getNames(rt)
# Read in a subset of branches (varsList), M rows:
df <- toR(rt, varsList, nEntries=M)

• Recently became aware of ROOTR, and I look forward to playing with that in the very near future

Getting and cleaning data in R

• data.table is extremely useful here:
  • fread found to be at least twice as fast as other methods I tried for importing my data
  • Helps me clean and filter my data and is super-fast, especially when using keys:

setkey(DT, numTracks) # Set number of particle tracks to be the key
DT <- DT[!.(1)] # Remove all single-track jets
DT[, (bad.cols) := NULL] # Remove junk columns

• digest is also useful for removing duplicate columns by fast comparison of hashes:

duplicate.columns <- names(DT)[duplicated(lapply(DT, digest))]
DT[, (duplicate.columns) := NULL]

• knitr and R Markdown used everywhere to document process; broke workflow into chunks, one R Markdown file for each, saving intermediate results along the way

More data munging

• To give me some extra space in RAM to work I used SOAR (stored object caches for R):

Sys.setenv(R_LOCAL_CACHE = "soar_cache")
Store(DT) # data.table now stored as RData file on disk and out of RAM

• caret also provides some useful data-munging; I could reduce the size of my data by more than 50% with a conservative cut on correlations between features:

highly.correlated <- findCorrelation(
  cor(DT[,-ncol(DT), with = FALSE], method = "pearson"),
  cutoff = 0.95, names = TRUE)

• Removing duplicate and highly correlated features was critical for enabling my data to fit in RAM
  • To preserve interpretability, I prefer to choose which features to retain instead of letting caret pick features that might have less explanatory value

Feature ranking & selection

• How should we find the features that provide the best discrimination between the processes or physics entities that we wish to classify?
• Given a newly-proposed discriminant variable, how can we rank this new variable against those we already know?
• We can use domain knowledge to drill down to what are believed to be the best discriminants; observables that:
  • Can explain most of the variance in the data
  • Are minimally correlated with each other
  • Provide the best predictive power
• How to optimally search the feature space? (Manual inspection may be impractical for a large feature set)

Summary

• It's often said that 80% of data analysis is spent on data munging\(^*\) \(-\) this has certainly been true in my case
• However, I've found a good set of tools for streamlining this process; I've shared what I found most useful here
• To the best of my knowledge, nobody has tried to do a particle physics analysis entirely in R before
  • Problems with large data, but workarounds exist
  • Insights gained will be valuable for helping me decide where to direct my efforts later when building a final model
• I didn't have to spend time writing a ton of code or worrying about dangling pointers, etc.
• R lets me focus on achieving the goals of my analysis

* Exploratory Data Mining and Data Cleaning, Dasu T, Johnson T (2003), John Wiley & Sons

Thanks!