Machine Learning Showdown: Apache Mahout vs. Weka | Hacker News Reader

Machine Learning Showdown: Apache Mahout vs. Weka

blog.algorithmia.com

54 pointsdoppenhe33 comments

Most data scientist these days use scikit-learn or R. Weka is really out of fashion. Mahout and mllib are difficult to use and perform less. Often it's better to just down-sample or rent an EC2 instance with a lot of memory.

Weka is definitely more old-school, but it has a LOT of algorithms available. Weka and Mahout are the two biggest ML libraries on the JVM, but we couldn't find any direct head-to-head comparison so this was the result. In the future we plan to also add scikit and mllib and more in the future.

Your point about being difficult to use is exactly the problem that Algorithmia solves.

This is almost apples and oranges. Mahout's power lies in it's ability to handle huge amounts of data, in a parallel fashion. Weka (which is rarely used these days anyways) is for smaller problems and experimentation.

None of these (Mahout and Weka) are mainstream anymore. For large-scale classification, people are using packages like VW[1] . And for small-scale experimentation, SciKit or R.

[1] http://hunch.net/~vw

IIRC VW doesn't have the easy integration with the hadoop ecosystem that mahout does, I am not sure everyone has moved to VW.

Doesn't need it. If you're computing a linear/logistic regression via gradient descent, performance-oriented C code on a single machine using local filesystem/caches will beat a hadoop-based algorithm for just about any size dataset.

I meant that shops are likely to already have data sitting in hdfs/hive/whatever which they can trivially use with mahout, while they may not have it sitting on a single file system on a beefy machine.

VW works on Hadoop just fine, though it's not very "clean" and native. But it works.

They can hadoop fs -get it or query it from hive.

Because you want to coalesce all your model updates for every pass over the data, the long startup-time for hadoop jobs actually plays into this. You can have hundreds of millions of samples, in a sparse space of hundreds of millions of parameters, and do it faster on a single node using VW than in a hadoop cluster of equivalent nodes running Mahout.

Rapidminer is the jam for prototyping ML processes. It's so powerfully useful I've always been surprised they've kept it free for so long (they have a pay version but it's not necessary). In addition to its own algorithms, it has a plugin that wraps Weka so you get all those if you want them too. I'm I n no way connected with them, just a big fan of it over every other ML library or tool I've seen. If I could buy stock I would...

for our HN friends direct invite https://algorithmia.com/signup?invite=HN24hr

algorithmia.com looks like a fantastic idea, never heard of it before. What do you use to wrangle your infrastructure?

Thanks! We use multiple cloud providers (Digital Ocean and AWS being our primary), Play Framework, AKKA , Ansible for deployment, Shippable for our CI and a good amount of custom clever code by our back end team.

Great idea, like a programmableweb.com but for core algorithms.

Thanks!

For ML, Spark MLlib is a solid choice.

For large scale, distributed stats I'd go with SparkR.

https://spark.apache.org/

we love Spark MLib

I'm not very experienced in Machine learning, just dabbled around a bit, so maybe someone could explain me this:

Looking at the graph number of trees vs accuracy, I would have expected that the line would asymptotically reach a maximum accuracy given more and more trees; however for weka it looks quite wavy and for mahout it even looks as if there's an optimum and more trees are worse.

Or is it just noise and I'm interpreting too much?

Usually RF will improve up to some point, and then the test (or OOB) accuracy will rattle around a little bit, centered around some final value. RF should never overfit from having too many trees. Both plots start at 50 trees, so it looks like the 'improve' part of that is over by then, and what you are seeing here is variation due to different trees being slightly better/worse. Incidentally, that means that the variation in both plots is probably not meaningful and that 'best' score of 99.4% at 250 trees is probably basically an outlier.

There is something bothering me about this...Weka's accuracy seems quite high in comparison to the results at Yann LeCun's MNIST page [1]. Its hard for me to believe that "the answer" to the MNIST problem is "use WEKA's RF". [1]http://yann.lecun.com/exdb/mnist/

Ugh... comparing random algorithms without showings error bounds on the accuracies.

the point here was more just to get both things running with a slick and more-or-less common interface. you're quite right that this isn't rigorous enough to settle the question between two algorithms in a domain where they both work somewhat - but stay tuned, rigorous analysis and comparison is high on our priority list.

I'm surprised that Spark's mllib wasn't included in this comparison.

next :) We started with these two because they are both available in Algorithmia now.

Just playing around with it, do typical strategies for using these tools include "bad" data? I drew a '-' and got '4' as the guess, which feels very wrong.

just a function of our simple demo, it returns confidence intervals for what it thinks it is of digits between 0-9 and we pick the top one. Not perfect by any means but shows how easily we can compare the libraries against each other.

WEKA ist really slow, at least when trying out the 'example'... And why choosing these two? There are a lot more. But anyway, thx for the comparison

We chose these two because they are the most common JAVA libraries used. We are adding new ones regularly to our library.

WEKA seems to take forever to classify a digit for their demo, also i wonder why there are various drops in performance when using 200 and 300 trees

traffic slowed down our system there for a second should be up and running at speed again.

works now! However it keeps mis-classifying my 4's.

http://imgur.com/mYRdIC0 http://imgur.com/cLQn74O http://imgur.com/bg2UaJN http://imgur.com/66oUGdM http://imgur.com/oU92V09 http://imgur.com/qvILooJ http://imgur.com/URyRiOB http://imgur.com/AybUI20 http://imgur.com/kugzG2D

IIRC the MINST data has a mix of style of 4's. I tried it with an open top, and it worked correctly. I wondering if they just happened to sample a majority of 4s that have an open top.

Example: http://imgur.com/mRRz1L3

Can anyone summarize the general workflow for using these analysis tools? Just looking for a high level intro and maybe a link to more detail.

I don't know anybody seriously using either of these. Mahout has bad implementations and Weka is showing it's age.

Reply on news.ycombinator.com