Define a "standard" process to follow when solving simple data science problems

[Drezil]

Some short things this should talk about:

• Data preparation

  • cleaning (removal of incomplete data, lookup of further data, get everything in one matrix/table/format)
  • whitening (normalize, center & de-corrolate - Warning: throws away corrolation-information!)
  • dimension-reduction (PCA, ICA, ... - Warning: throws away information!)

• Algorithm selection

  • supervised? unsupervised?
  • typical solutions for typical problems (classification, corrolation, non-metric-solutions (i.e. NLP with suffix-trees, edit-distance, etc.))
  • for each algorithm
    • when and when NOT to use
    • further reading

• Ways to present/interpret results

  • statistical significance?
  • typical tests/metrics (AUC, F_1 score, sensitivity/specificity, etc.)

[ocramz]

@ixxie wants to write about reproducibility with Jupyter and Nix, I've added to the DH members'list, he should see this soon as well

[ixxie]

Hmmm, I am not sure if this is quite relevant to this; my goal is more to try and create easily reproducible infrastructure as code, i.e. to allow anybody to deploy a data science platform relatively easily. Reproducibility of individual computations is also of great interest and Nix can help with this, but I don't know much about this atm (would be willing to look into it some time!).

FWIW, it seems a bit far fetched to be able to specify a simple decision tree recipe for doing data science; the way I would approach this is to think of it like a bipartite graph: list some problems (e.g. tokenization, classification, clustering, etc) and some algorithms (CRFs, RNNs, HDBSCAN) and link between them.