begriffs

A Survey of Data Science

November 30, 2014

What does it take to be an outstanding data scientist? I decided to find out and reviewed syllabi from every data science bootcamp, the requirements for dozens of data jobs, and scores of slideshares and wikipedia articles. I think I’ve identified and classified the core skills.

Fundamentally there are six abilities of a competent data scientist.

  • acquiring data in a highly-available distributed pipeline
  • cleaning it of duplication and errors
  • enhancing it with knowledge gained from unsupervised learning and nlp
  • inferring properties of a population from a sample, and finding relations of statistical variables
  • predicting future events through supervised learning and statistical models
  • communicating my findings to people and external programs

As many articles mentioned, dealing with data is a team job and no single individual is expected to know everything. There are two distinct specializations though, data science and engineering. The latter deails with pipelines, wrangling, availability, and cluster management. The former with asking the right questions, designing experiments, and statistical analysis.

I compiled a list of specific technologies and techniques and fit them into the sixfold classification of abilities. These appear to be the core skills of the profession.

  • Acquire
    • workers and other pieces are managed by mesos
    • web scrapers and other events as kafka producers
    • mesos runs them in docker containers with the marathon framework
    • feeds into a “lambda architecture”
      • the master dataset accumulates immutably in hdfs
        • encode the data with the thrift format to add a schema
        • thence through hadoop jobs to make precomputed views
        • store views in elephantdb and recalculate periodically
        • client access is read-only from elephant
      • data forks into storm for real time “speed layer” processing
        • which redundantly calculates the views we are doing on hadoop
        • the results go into cassandra which is read/write
        • the realtime layer relies on incremental algorithms to update the state in that database
  • Clean
    • record linking and deduping
      • soundex
      • jaro–winkler distance
      • unicode normalization
    • outlier identification
      • normality testing
      • Grubbs’ test for outliers (given normality)
    • correcting errors with mechanical turk
    • create a derived hdfs dataset, the “golden master”
  • Enhance
    • feature extraction
      • principal component analysis
      • k-NN
    • unsupervised clustering
      • k-means and finding k with the gap statistic
      • hierarchical (agglomerative, divisive)
      • density (DBSCAN)
      • bagging
    • nlp
      • stemming
      • bag of words
      • tf–idf
      • topic models
  • Infer
    • point estimation
      • maximum likelihood
    • interval estimation
      • credible intervals
    • survival analysis
      • kaplan-meier estimator
    • hypothesis testing
      • chi squared
      • t-test
    • correlation
  • Predict
    • classification
      • logistic regression
      • decision trees and random forests
      • naive bayes
      • variance-bias decomposition
    • recommendations
      • collaborative filtering
      • content filtering
  • Communicate
    • descriptive statistics
      • measures of central tendency
      • measures of dispersion
    • visualization
      • d3 (vega, rickshaw)
      • realtime dashboard
    • exposing results in restful api
      • export to postgres with sqoop
      • generate api with dbapi
      • document with raml

This seems like the state of the art although there is some dissenting opinion such as Questioning the Lambda Architecture. Ultimately I think the engineering side will be moving to a purely functional paradigm. Perhaps it could be written in Haskell, using libraries like hailstorm, haskakafka, and hadron. Everything I read agreed that Hadoop and associated Java tools are creaky beasts that require a lot of fiddling.

Although the engineering side looks exhilerating (unleashing hordes of data scrapers and coordinating a big pipeline), the science side may be a calmer career bet with its longer-lasting knowledge. However even there I discovered some fundamental disagreements between Frequentists and Bayesians. Either way it’s all fascinating stuff and I’d like to create some test compute clusters to up my game.