I have tried to use MeV26, Bayesia software and R for making Bayesian network from 26 Columns of gene expression microarray numbers (.csv file, 652 genes there). Does anybody experienced can advise what software and scripts to use and what books and tutorials are best for that task? Are there any Python or Ruby libraries for that?
Thank you
Software Tools:
The easiest way would be to use WEKA. Simply import your data into WEKA, select Bayesian/ Bayesian Network (BN) as your classifier option, learn a structure and look at your classification performance.
The second would be to use R with the bnlearn package.
In general, you can find BN libraries in all major languages. I am not familiar with Ruby but Python has this.
But again, I would advise first using WEKA, as it will give you results almost instantaneously, which you can use later to benchmark your more detailed results that you obtain by getting your hands dirty / coding in whatever language.
Reading:
Obviously, there are many articles published on BNs but you may not have access to them and I presume you don't want to pay to buy a book straight-away. MatLab BNT's developer K. Murphy has a nice introductory article. Furthermore the BNT's manual itself provides a nice, brief and hands-on (if you have MatLab) introduction to the use of BNs.
Given the stress toward reproducible science, I was wondering if my recent work warrants the inclusion of example code in the publication. The datasets that I am using are quite big, so it wouldn't make sense to publish those necessarity - However, the statistical methods that I apply within R are not generally known to my audience (although I would think that they should be).
I'm using empirical orthogonal function analysis (EOF) and generalized additive models (GAM) within my analysis. GAM, in particular, is widely used in ecological studies, but less so within the physical sciences - my work spans both disciplines.
I definitely refer to the R packages that I use, and it wouldn't really be difficult for a reviewer / reader to look for those references (and included examples) themselves. So, my question is, what situations are most appropriate for the inclusion of reproducible code in a publication?
Code is the most accurate representation of what you actually did. Therefore, in my view you should always aim to publish code alongside your article.
However, editor resistance to this is pretty strong. The fear is that if the reviewer had access to the code, then the journal looks pretty bad if a substantive coding mistake is later found. This is not a hypothetical fear, given the Levitt paper, etc.
Knuth has some strong views on literate programming that you should be able to cite as justification. If you can't convince the journal to accept your code as an integral piece of the publication, consider publishing it on your personal website (the approach taken by e.g. Raj Chetty for many of his papers) or publish it as an R package.
Finally, here's a note I wrote to my programming students:
Consider publishing your code. Doing so will act as a commitment
device which will encourage good habits--habits that make your own
work easier. Publishing your code also makes it easier for others to
extend your analysis, which can result in more citations of your work.
Releasing your code is good academic practice as well: it is the
truest testament to your analysis. And offering your program to the
world shows off the beautiful coding skills which you are about to
acquire.
A basic tenet of science is reproducibility. So the answer would be to "include" code required to conduct your analysis to every paper/publication that is based on data analysis.
I say "include" because you don't need to put the R code directly into the paper. Many if not most journals allow supplementary material which is an option. Alternative, supply your script to one of the many Science data archiving sites (Such as Figshare) and then (and here is the killer!) cite your own script using the DOI that Figshare gives to your deposited script. If you can post the data too, then all the better; Figshare doesn't really care too much about big data sets.
The above applies to code where you are using other packages and your R script does things like loads and formats data, calls functions from other packages and then plots or displays output/results. If you have developed new R code to implement a particular method then I would say package the code as an R package and submit that to CRAN or r-forge or something like that.
From your description, the former (deposit the analysis script in a repo) would be most appropriate.
We recently had a discussion at our research institute regarding reproducible research. The incentive came from the Nature editorial (http://arstechnica.com/science/2012/02/science-code-should-be-open-source-according-to-editorial/) which argued that all your code should be published. I whole heartedly agree with this. Even though your dataset is very big, publishing the R code that you used to create your results makes it crystal clear what you did. Often times the methods of a paper do not contain sufficient detail to reproduce the result, the code is quite a help in this case.
Crossposted with STATS.se since this problem could straddle both STATs.se/SO
https://stats.stackexchange.com/questions/17712/parallelize-solve-for-ax-b
I have some extremely large sparse matrices created using spMatrix function from the matrix package.
Using the solve() function works for my Ax=b issue, but it takes a very long time. Several days.
I noticed that http://cran.r-project.org/web/packages/RScaLAPACK/RScaLAPACK.pdf
appears to have a function that can parallelize the solve function, however, it can take several weeks to get new packages installed on this particular server.
The server already has the snow package installed it.
So
Is there a way of using snow to parallelize this operation?
If not, are there other ways to speed up this type of operation?
Are there other packages like RScaLAPACK? My search on RScaLAPACK seemed to suggest people had a lot of issues with it.
Thanks.
[EDIT] -- Additional details
The matrices are about 370,000 x 370,000.
I'm using it to solve for alpha centrality, http://en.wikipedia.org/wiki/Alpha_centrality. I was originally using the alpha centrality function in the igraph package, but it would crash R.
More details
This is on a single machine with 12 cores and 96 gigs of memory (I believe)
It's a directed graph along the lines of paper citation relationships.
Calculating condition number and density will take awhile. Will post as it comes available.
Will crosspost on stat.SE and will add a link back to here
[Update 1: For those just tuning in: The original question involved parallelizing computations to solving a regression problem; given that the underlying problem is related to alpha centrality, some of the issues, such as bagging and regularized regression may not be as immediately applicable, though that leads down the path of further statistical discussions.]
There are a bundle of issues to address here, from the infrastructural to the statistical.
Infrastructure
[Updated - also see Update #2 below.]
Regarding parallelized linear solvers, you can replace R's BLAS / LAPACK library with one that supports multithreaded computations, such as ATLAS, Goto BLAS, Intel's MKL, or AMD's ACML. Personally, I use AMD's version. ATLAS is irritating, because one fixes the number of cores at compilation, not at run-time. MKL is commercial. Goto is not well supported anymore, but is often the fastest, but only by a slight margin. It's under the BSD license. You can also look at Revolution Analytics's R, which includes, I think, the Intel libraries.
So, you can start using all of the cores right away, with a simple back-end change. This could give you a 12X speedup (b/c of the number of cores) or potentially much more (b/c of better implementation). If that brings down the time to an acceptable range, then you're done. :) But, changing the statistical methods could be even better.
You've not mentioned the amount of RAM available (or the distribution of it per core or machine), but A sparse solver should be pretty smart about managing RAM accesses and not try to chew on too much data at once. Nonetheless, if it is on one machine and if things are being done naively, then you may encounter a lot of swapping. In that case, take a look at packages like biglm, bigmemory, ff, and others. The former addresses solving linear equations (or GLMs, rather) in limited memory, the latter two address shared memory (i.e. memory mapping and file-based storage), which is handy for very large objects. More packages (e.g. speedglm and others) can be found at the CRAN Task View for HPC.
A semi-statistical, semi-computational issue is to address visualization of your matrix. Try sorting by the support per row & column (identical if graph is undirected, else do one then the other, or try a reordering method like reverse Cuthill-McKee), and then use image() to plot the matrix. It would be interesting to see how this is shaped, and that affects which computational and statistical methods one could try.
Another suggestion: Can you migrate to Amazon's EC2? It is inexpensive, and you can manage your own installation. If nothing else, you can prototype what you need and migrate it in-house once you have tested the speedups. JD Long has a package called segue that apparently makes life easier for distributing jobs on Amazon's Elastic MapReduce infrastructure. No need to migrate to EC2 if you have 96GB of RAM and 12 cores - distributing it could speed things up, but that's not the issue here. Just getting 100% utilization on this machine would be a good improvement.
Statistical
Next up are multiple simple statistical issues:
BAGGING You could consider sampling subsets of your data in order to fit the models and then bag your models. This can give you a speedup. This can allow you to distribute your computations on as many machines & cores as you have available. You can use SNOW, along with foreach.
REGULARIZATION The glmnet supports sparse matrices and is very fast. You would be wise to test it out. Be careful about ill-conditioned matrices and very small values of lambda.
RANK Your matrices are sparse: are they full rank? If they are not, that could be part of the issue you're facing. When matrices are either singular or very nearly so (check your estimated condition number, or at least look at how your 1st and Nth eigenvalues compare - if there's a steep drop off, you're in trouble - you might check eval1 versus ev2,...,ev10,...). Again, if you have nearly singular matrices, then you need to go back to something like glmnet to shrink out the variables are either collinear or have very low support.
BOUNDING Can you reduce the bandwidth of your matrix? If you can block diagonalize it, that's great, but you'll likely have cliques and members of multiple cliques. If you can trim the most poorly connected members, then you may be able to estimate their alpha centrality as being upper bounded by the lowest value in the same clique. There are some packages in R that are good for this sort of thing (check out Reverse Cuthill-McKee; or simply look to see how you'd convert it into rectangles, often relating to cliques or much smaller groups). If you have multiple disconnected components, then, by all means, separate the data into separate matrices.
ALTERNATIVES Are you wedded to the Alpha Centrality? There may be other measures that are monotonically correlated (i.e. have high rank correlation) with the same value that could be calculated more cheaply or at least implemented quite efficiently. If those will work, then your analyses could proceed with a lot less effort. I have a few ideas, but SO isn't really the place to go about that discussion.
For more statistical perspectives, appropriate Q&A should occur on the stats.stackexchange.com, Cross-Validated.
Update 2: I was a bit too quick in answering and didn't address this from the long-term perspective. If you are planning to do research on such systems for the long-term, you should look at other solvers that may be more applicable to your type of data and computing infrastructure. Here is a very nice directory of the options for both solvers and pre-conditioners. It seems this doesn't include IBM's "Watson" solver suite. Although it may take weeks to get software installed, it's quite possible that one of the packages is already installed if you have a good HPC administrator.
Also, keep in mind that R packages can be installed to the user directory - you need not have a package installed in the general directory. If you need to execute something as a user other than yourself, you could also download a package to the scratch or temporary space (if you're running within just 1 R instance, but using multiple cores, check out tempdir).
I'm looking to port our home-grown platform of various machine learning algorithms from C# to a more robust data mining platform such as R. While it's obvious R is great at many types of data mining tasks, it is not clear to me if it can be used for text classification.
Specifically, we extract a list of bigrams from the text and then classify it into one of 15 different categories, eg:
Bigram list: jewelry, books, watches, shoes, department store
-> Category: Shopping
We'd want to both train the models in R as well as hook up to a database to perform this on a larger scale.
Can it be done in R?
Hmm, I am rather starting to look into Machine Learning, but I might have a suggestion: have you considered Weka? There's a bunch of various algorithms around and there'S IS some documentation. Plus, there is an R package RWeka that makes use of the Weka jars.
EDIT:
There is also a nice, comprehensive read by Witten et al. : Data mining that contains an extensive description of Weka among other interesting things. Look into the API opportunities.
I will be analysing vast amount of network traffic related data shortly, and will pre-process the data in order to analyse it. I have found that R and SPSS are among the most popular tools for statistical analysis. I will also be generating quite a lot of graphs and charts. Therefore, I was wondering what is the basic difference between these two softwares.
I am not asking which one is better, but just wanted to know what are the difference in terms of workflow between the two (besides the fact that SPSS has a GUI). I will be mostly working with scripts in either case anyway so I wanted to know about the other differences.
Here is something that I posted to the R-help mailing list a while back, but I think that it gives a good high level overview of the general difference in R and SPSS:
When talking about user friendlyness
of computer software I like the
analogy of cars vs. busses:
Busses are very easy to use, you just
need to know which bus to get on,
where to get on, and where to get off
(and you need to pay your fare). Cars
on the other hand require much more
work, you need to have some type of
map or directions (even if the map is
in your head), you need to put gas in
every now and then, you need to know
the rules of the road (have some type
of drivers licence). The big advantage
of the car is that it can take you a
bunch of places that the bus does not
go and it is quicker for some trips
that would require transfering between
busses.
Using this analogy programs like SPSS
are busses, easy to use for the
standard things, but very frustrating
if you want to do something that is
not already preprogrammed.
R is a 4-wheel drive SUV (though
environmentally friendly) with a bike
on the back, a kayak on top, good
walking and running shoes in the
pasenger seat, and mountain climbing
and spelunking gear in the back.
R can take you anywhere you want to go
if you take time to leard how to use
the equipment, but that is going to
take longer than learning where the
bus stops are in SPSS.
There are GUIs for R that make it a bit easier to use, but also limit the functionality that can be used that easily. SPSS does have scripting which takes it beyond being a mere bus, but the general phylosophy of SPSS steers people towards the GUI rather than the scripts.
I work at a company that uses SPSS for the majority of our data analysis, and for a variety of reasons - I have started trying to use R for more and more of my own analysis. Some of the biggest differences I have run into include:
Output of tables - SPSS has basic tables, general tables, custom tables, etc that are all output to that nifty data viewer or whatever they call it. These can relatively easily be transported to Word Documents or Excel sheets for further analysis / presentation. The equivalent function in R involves learning LaTex or using a odfWeave or Lyx or something of that nature.
Labeling of data --> SPSS does a pretty good job with the variable labels and value labels. I haven't found a robust solution for R to accomplish this same task.
You mention that you are going to be scripting most of your work, and personally I find SPSS's scripting syntax absolutely horrendous, to the point that I've stopped working with SPSS whenever possible. R syntax seems much more logical and follows programming standards more closely AND there is a very active community to rely on should you run into trouble (SO for instance). I haven't found a good SPSS community to ask questions of when I run into problems.
Others have pointed out some of the big differences in terms of cost and functionality of the programs. If you have to collaborate with others, their comfort level with SPSS or R should play a factor as you don't want to be the only one in your group that can work on or edit a script that you wrote in the future.
If you are going to be learning R, this post on the stats exchange website has a bunch of great resources for learning R: https://stats.stackexchange.com/questions/138/resources-for-learning-r
The initial workflow for SPSS involves justifying writing a big fat cheque. R is freely available.
R has a single language for 'scripting', but don't think of it like that, R is really a programming language with great data manipulation, statistics, and graphics functionality built in. SPSS has 'Syntax', 'Scripts' and is also scriptable in Python.
Another biggie is that SPSS squeezes its data into a spreadsheety table structure. Dealing with other data structures is probably very hard, but comes naturally to R. I wouldn't know where to start handling network graph type data in SPSS, but there's a package to do it for R.
Also with R you can integrate your workflow with your reporting by using Sweave - you write a document with embedded bits of R code that generate plots or tables, run the file through the system and out comes the report as a PDF. Great for when you want to do a weekly report, or you do a body of work and then the boss gives you an updated data set. Re-run, read it over, its done.
But you know, your call...
Well, are you a decent programmer? If you are, then it's worthwhile to learn R. You can do more with your data, both in terms of manipulation and statistical modeling, than you can with SPSS, and your graphs will likely be better too. On the other hand, if you've never really programmed before, or find the idea of spending several months becoming a programmer intimidating, you'll probably get more value out of SPSS. The level of stuff that you can do with R without diving into its power as a full-fledged programming language probably doesn't justify the effort.
There's another option -- collaborate. Do you know someone you can work with on your project (you don't say whether it's academic or industry, but either way...), who knows R well?
There's an interesting (and reasonably fair) comparison between a number of stats tools here
http://anyall.org/blog/2009/02/comparison-of-data-analysis-packages-r-matlab-scipy-excel-sas-spss-stata/
I work with both in a company and can say the following:
If you have a large team of different people (not all data scientists), SPSS is useful because it is plain (relatively) to understand. For example, if users are going to run a model to get an output (sales estimates, etc), SPSS is clear and easy to use.
That said, I find R better in almost every other sense:
R is faster (although, sometimes debatable)
As stated previously, the syntax in SPSS is aweful (I can't stress this enough). On the other hand, R can be painful to learn, but there are tons of resources online and in the end it pays much more because of the different things you can do.
Again, like everyone else says, the sky is the limit with R. Tons of packages, resources and more importantly: indepedence to do as you please. In my organization we have some very high level functions that get a lot done. The hard part is creating them once, but then they perform complicated tasks that SPSS would tangle in a never ending web of canvas. This is specially true for things like loops.
It is often overlooked, but R also has plenty of features to cooperate between teams (github integration with RStudio, and easy package building with devtools).
Actually, if everyone in your organization knows R, all you need is to maintain a basic package on github to share everything. This of course is not the norm, which is why I think SPSS, although a worst product, still has a market.
I have not data for it, but from my experience I can tell you one thing:
SPSS is a lot slower than R. (And with a lot, I really mean a lot)
The magnitude of the difference is probably as big as the one between C++ and R.
For example, I never have to wait longer than a couple of seconds in R. Using SPSS and similar data, I had calculations that took longer than 10 minutes.
As an unrelated side note: In my eyes, in the recent discussion on the speed of R, this point was somehow overlooked (i.e., the comparison with SPSS). Furthermore, I am astonished how this discussion popped up for a while and silently disappeared again.
There are some great responses above, but I will try to provide my 2 cents. My department completely relies on SPSS for our work, but in recent months, I have been making a conscious effort to learn R; in part, for some of the reasons itemized above (speed, vast data structures, available packages, etc.)
That said, here are a few things I have picked up along the way:
Unless you have some experience programming, I think creating summary tables in CTABLES destroys any available option in R. To date, I am unaware package that can replicate what can be created using Custom Tables.
SPSS does appear to be slower when scripting, and yes, SPSS syntax is terrible. That said, I have found that scipts in SPSS can always be improved but using the EXECUTE command sparingly.
SPSS and R can interface with each other, although it appears that it's one way (only when using R inside of SPSS, not the other way around). That said, I have found this to be of little use other than if I want to use ggplot2 or for some other advanced data management techniques. (I despise SPSS macros).
I have long felt that "reporting" work created in SPSS is far inferior to other solutions. As mentioned above, if you can leverage LaTex and Sweave, you will be very happy with your efficient workflows.
I have been able to do some advanced analysis by leveraging OMS in SPSS. Almost everything can be routed to a new dataset, but I have found that most SPSS users don't use this functionality. Also, when looking at examples in R, it just feels "easier" than using OMS.
In short, I find myself using SPSS when I can't figure it out quickly in R, but I sincerely have every intention of getting away from SPSS and using R entirely at some point in the near future.
SPSS provides a GUI to easily integrate existing R programs or develop new ones. For more info, see the SPSS Community on IBM Developer Works.
#Henrik, I did the same task you have mentioned (C++ and R) on SPSS. And it turned out that SPSS is faster compared to R on this one. In my case SPSS is aprox. 7 times faster. I am surprised about it.
Here is a code I used in SPSS.
data list free
/x (f8.3).
begin data
1
end data.
comp n = 1e6.
comp t1 = $time.
loop #rep = 1 to 10.
comp x = 1.
loop #i=1 to n.
comp x = 1/(1+x).
end loop.
end loop.
comp t2 = $time.
comp elipsed = t2 - t1.
form elipsed (f8.2).
exe.
Check out this video why is good to combine SPSS and R...
Link
http://bluemixanalytics.wordpress.com/2014/08/29/7-good-reasons-to-combine-ibm-spss-analytics-and-r/
If you have a compatible copy of R installed, you can connect to it from IBM SPSS Modeler and carry out model building and model scoring using custom R algorithms that can be deployed in IBM SPSS Modeler. You must also have a copy of IBM SPSS Modeler - Essentials for R installed. IBM SPSS Modeler - Essentials for R provides you with tools you need to start developing custom R applications for use with IBM SPSS Modeler.
The truth is: both packages are useful if you do data analysis professionally. Sure, R / RStudio has more statistical methods implemented than SPSS. But SPSS is much easier to use and gives more information per each button click. And, therefore, it is faster to exploit whenever a particular analysis is implemented in both R and SPSS.
In the modern age, neither CPU nor memory is the most valuable resource. Researcher's time is the most valuable resource. Also, tables in SPSS are more visually pleasing, in my opinion.
In summary, R and SPSS complement each other well.