>From my perspective, the state of the art of machine learning is with systems like Tensorflow and dl4j. If you can deal with the limits of a non-clustered GPU system, then Theano and Cafe are very useful. Keras papers over the difference between different back-ends nicely.
Tensorflow and Theano can do a lot of mathematical and linear (tensor, actually) algebra work nicely, especially if there is an optimization problem lurking. NVidia also has a very strong commercial offering that supports their GPU clustering well. Spark ML lags this state of the art very far behind, but is still useful for simpler situations. For recommendations, the situation is very different. Almost all applications are most easily and often most accurately solved using an indicator-based approach and the go-to implementation of this is Mahout. There is a lot of noise in the world about factorization-based recommendation using ALS and such, but the noise is not warranted. Deploying a recommender in a search engine is just better. I have not personally used Samsara much, but the idea of a strong optimizer over the top of a nice syntax for linear algebra is a good one. On Tue, Jan 31, 2017 at 9:21 AM, Pat Ferrel <p...@occamsmachete.com> wrote: > My perspective comes from the data side. I work in recommenders and that > means log analysis for huge amounts of data. Even a small shop doing this > will immediately run our of the capacity in Python or R on a single node. > MLlib is a set of prepackaged algorithms that will work (mostly) with big > data. Mahout Samsara is the only general linear algebra tool I know of that > will natively let you interactively run R-like code on any size cluster, > then polish it for production all without changing tools, or language. > > Going from analytics to recommenders means a jump in data size of several > orders of magnitude and this is just one example. > > > On Jan 31, 2017, at 6:50 AM, Trevor Grant <trevor.d.gr...@gmail.com> > wrote: > > Hello Isabel and Florent, > > I'm currently working on a side-by-side demo of R / Python / SparkML(Mllib) > / Mahout, but in very broad strokes here is how I would compare them: > > R- Most statistical functionality. Most flexibility. Implement your own > algorithms- mathematically expressive language. Worst performance- handles > only "small" data sets. Language is 'math centric'. Easy to extend / > create new algos > > Python (sklearn/scikit) - Some mathematical / statistical functionality, > more focused on machine learning. Machine learning library very > sophisticated though. Much better performance than R, still only single > node. "small to medium" data sets. Language is 'programmer centric'. > Somewhat difficult to extend / create new algos > > SparkML / Mllib - Very Limited Mathematical functionality (usually collects > to driver to do anything of substance). Machine learning rudimentary > compared to sklearn, but still non-trivial one of the best available. > Exceeding performance, well suited to "big" data sets. Language is > 'programmer centric'. Very difficult to extend / create new algos. > > (FlinkML - Fits in same spot as SparkML, but significantly less developed) > > Mahout - Good mathematical functionality. Good performance relative to > underlying engine (possibly superior with MAHOUT-1885). Language is 'math > centric'. Well suited to "medium and big" data sets. Fairly easy to extend > / create new algos (MAHOUT-1856) > > I hope that provides a high level comparison. > > Re use cases- the tool to use depends on the job at hand. > Highly advanced mathematical model, small dataset or sampling from full > dataset OK -> Use R > Machine learning on small to medium data set or sampling from full dataset > OK -> Use Python / sklearn > Less sophisticated machine learning on Large dataset -> SparkML > Custom mathematical/statistical model on medium to large data -> Mahout > > ^^ All of this is just my opinion. > > Re: integration- > > We're working on that too. Recently MAHOUT-1896 added convenience methods > for interacting with MLLib type RDDs, and DataFrames > https://issues.apache.org/jira/browse/MAHOUT-1896 > > (No support yet for SparkML type dataframes, or spitting DRMs back out into > RDDs/DataFrames). > > Finally Docs: There has been some talk for sometime of migrating the > website from CMS to Jekyll and its something I strongly support. The CMS > makes it difficult to keep up with documentation, and Jekyll would open up > documentation /website maintenance to contributors. > > Trevor Grant > Data Scientist > https://github.com/rawkintrevo > http://stackexchange.com/users/3002022/rawkintrevo > http://trevorgrant.org > > *"Fortunate is he, who is able to know the causes of things." -Virgil* > > > On Tue, Jan 31, 2017 at 5:31 AM, Florent Empis <florent.em...@gmail.com> > wrote: > > > Hi, > > > > I am in the same spot as Isabel. > > Used to use/understand most of the «old» standalone mahout, now doing > some > > data transformation with spark, but I am not sure where Samsara fits in > the > > ecosystem. > > We also do quite a bit of computation in R. > > Basically we are willing to learn and support the project by for instance > > buying the books Rob mentioned, but a short doc with the outline Isabel > > describes would be great! > > > > Many thanks, > > > > Florent > > > > > > Le 31 janv. 2017 12:01, "Isabel Drost-Fromm" <isa...@apache.org> a > écrit : > > > > > > Hi, > > > > On Fri, Sep 16, 2016 at 11:36:03PM -0700, Andrew Musselman wrote: > >> and we're thinking about just how many pre-built algorithms we > >> should include in the library versus working on performance behind the > >> scenes. > > > > To pick this question up: I've been watching Mahout from a distance for > > quite > > some time. So from what limited background I have of Samsara I really > like > > it's > > approach to be able to run on more than one execution engine. > > > > To give some advise to downstream users in the field - what would be your > > advise > > for people tasked with concrete use cases (stuff like fraud detection, > > anomaly > > detection, learning search ranking functions, building a recommender > > system)? Is > > that something that can still be done with Mahout? What would it take to > > get > > from raw data to finished system? Is there something we can do to help > > users get > > that accomplished? Is there even interest from users in such a use case > > based > > perspective? If so, would there be interest among the Mahout committers > to > > help > > users publicly create docs/examples/modules to support these use cases? > > > > > > Isabel > > > >
