- 317comments
- —discuss
- 90comments
- 113comments
- 200comments
- 35comments
- 2comments
- 351comments
- 34comments
- 533comments
- 140comments
- 24comments
- 318comments
- 390comments
- 4comments
- 9comments
- 119comments
- 42comments
- 34comments
- 599comments
- 12comments
- 20comments
- 16comments
- 122comments
- 20comments
- 20comments
- 39comments
- 45comments
- —discuss
- 11comments
Let me give a concrete example. For the past year, I have been selling stuff on the interwebs through two payment processors one of them being PayPal.
The selling process was put together with a bunch of SaaS hooking everything together through webhooks and notifications.
Now I need to step it that control and produce a proper flow to handle sign up, subscription and payment.
Before doing that I'm analyzing and trying to conciliate all transactions to make sure the books are OK and nothing went unseen. There lies the problem. I have data coming from different sources such as databases, excel files, CSV exports and some JSON files.
At first, I started dealing with it by having all the data in CSV files and trying to make sense of them using code and running queries within the code.
As I found holes in the data I had to dig up more data from different sources and it became a pain to continue with code. I now imported everything into Postgres and have been "debugging" with SQL.
As I advanced through the process I had to generate a lot of routines to collect and match data. I also have to keep all the data files around and organized which is very hard to do because I'm all over the place trying to find where the problem is.
How do you handle with it? What kind of workflow? Any best practices or recommendations from people who do this for a living?
Pachyderm may be basically what you're looking for. It does data version control with/for language-agnostic pipelines that don't need to always redo the ETL phase. https://www.pachyderm.io
Dask-ML works with {scikit-learn, xgboost, tensorflow, TPOT,}. ETL is your responsibility. Loading things into parquet format affords a lot of flexibility in terms of (non-SQL) datastores or just efficiently packed files on disk that need to be paged into/over in RAM. http://ml.dask.org/examples/scale-scikit-learn.html
Sklearn.pipeline.Pipeline API: {fit(), transform(), predict(), score(),} https://scikit-learn.org/stable/modules/generated/sklearn.pi...
https://docs.featuretools.com can also minimize ad-hoc boilerplate ETL / feature engineering :
The PLoS 10 Simple Rules papers distill a number of best practices:
"Ten Simple Rules for Reproducible Computational Research" http://www.ploscompbiol.org/article/info%3Adoi%2F10.1371%2Fj...
“Ten Simple Rules for Creating a Good Data Management Plan” http://journals.plos.org/ploscompbiol/article?id=10.1371/jou...
In terms of the scientific method, a null hypothesis like "there is no significant relation between the [independent and dependent] variables" may be dangerously unprofessional p-hacking and data dredging; and may result in an overfit model that seems to predict or classify the training and test data (when split with e.g. sklearn.model_selection.train_test_split and a given random seed).
One of these days (in the happy new year!) I'll get around to updating these notes with the aforementioned tools and docs: https://wrdrd.github.io/docs/consulting/data-science#scienti...
IDK what https://kaggle.com/learn has specifically in terms of analysis workflow? Their docker containers have very many tools configured in a reproducible way: https://github.com/Kaggle/docker-python/blob/master/Dockerfi...