My friend Szilard is the author of this benchmark. He was attempting to answer the questions in the comments, but the responses did not show up. I am posting hi

My friend Szilard is the author of this benchmark. He was attempting to answer the questions in the comments, but the responses did not show up. I am posting his answers here for him: ~~~~ I was trying to answer each question, but got blocked, so here are some answers in one comment: 1. Yes, it's WIP. 2. Did RF+GBM so far, did not even start DL, but stay tuned... 3. For Spark I used number of partitions = number of cores 4. For why Spark is less accurate for RF see Databricks comments here: http://datascience.la/benchmarking-random-forest-implementations/#comment-53599 http://datascience.la/benchmarking-random-forest-implementat... (essentially current version aggregates votes, while it should aggregate probabilities). 5. The issue of Spark being slow is not only an overhead on small data, but lots of serializing/deserializing, garbage collection etc. 6. Non-linear SVMs scale badly, but one could look at what max size can e.g. svmlight do (results/pull requests are welcome) 7. I think Logistic regression in Python scikit-learn can be much improved (as I mentioned on github) by using a sparse format (someone should do it) 8. The difference between implementations in somewhat surprising, but some of the tools take different approaches (e.g. RF in Python/xgboost works with the data, in H2O/Spark it builds bins from the data etc.) There are tricks, options available only in 1 or 2 etc. I was trying to do my best to match parameter values/setup, but it's not the same. I also wrote a blog post (a slightly more organized text for RF than the github README): http://datascience.la/benchmarking-random-forest-implementations/ http://datascience.la/benchmarking-random-forest-implementat... There is also a recording of a talk I gave at a Machine Learning meetup if you want more insights: https://www.youtube.com/watch?v=DK87lCLH_6A https://www.youtube.com/watch?v=DK87lCLH_6A