christopheraden

Because normalization reduces the influence of variables that have a higher variance. In raw data, if you have marathon times and heights for runners in a race, and you measure the times in minutes and the heights in inches, the influence on the principal components from the times will likely be much more.

Like all problems in statistics, it ought to depend on the specific task at hand. If there is some a priori reason to use the original scale (or a different re-weighting), it ought to be used. In general, PCA on correlation matrices is much preferred for exactly the reason you mention.

It is somewhat similar. The procedure I describe is normalizing (using z-scores instead of raw data) the data. The difference, as far as I can tell, is that normalizing retains the correlation structure of the data, while whitening uncorrelates the observations. R is the lingua franca of the academic statisticians, so you might not derive a huge amount of value from this, but this question was asked a couple years ago on Stats.SE: http://stats.stackexchange.com/questions/53/pca-on-correlati...

Yes, of course you will--because you've normalized the data. If you had run PCA on just ingredients instead of on the normalized ingredients, I would imagine the results would be different. I can't verify since I don't have matlab on this computer, but doing PCA on raw data with one set of units will produce a different PCA doing it on data with another set of units.

While I agree with you, if the author believed the entire method is invalid in all situations, he could make a strong argument that it wasn't worth the space to put it in at all. It would be a long book if it was filled with all the methods currently in production that did _not_ work!

There is actually a fix to this problem in classical statistics (as far back as Pearson in the early 20th century) if the object is to perform PCA on data matrices: Don't use the covariance matrix for PCA. The issue of units is only a problem if you're using a matrix that has units itself. This problem is readily solved by using the correlation matrix instead, which is dimensionless by definition. The downside to this circumvention is that you have essentially re-weighted each of your variables, so the weight contributed by each variable is more similar. This may not be what you want.

If PCA is not being used to reduce the dimensionality of multivariate data, this might be invalid. There are other uses of PCA besides working on data (that image reduction technique using SVD comes to mind) that he might be addressing.

If you want a treatment of PCA in a respected text (at least by the statistical community--not sure what the ML people think ;) ), look no further than Hastie, Tibshirani and Friedman's Elements of Stastical Learning.

http://www.stanford.edu/~hastie/local.ftp/Springer/OLD//ESLI...

A great point, but therein lies my biggest complaint with the simplification of statistics that I see in the startup world--sometimes the technical details are actually important. As an analogy, while mass-production has given us a car that anyone can operate, we are largely helpless when one breaks down. Complications abound when individuals try to leverage an overly-simplistic view of a subject (raise your hand if you've heard "We are 95% sure the true [...] lies in this confidence interval").

To the credit of the shadier individuals in my profession, this histogram subtlety nicely highlights how it can be quite easy to bend the data to your argument using ad-hoc procedures (KDEs, hists, QQs, boxplots). A carefully chosen bin width, smoothing parameter, or covariate can present a different view of the data than some other parameter/covariate. That's why it's nice to have other statisticians capable of reproducing and disseminating the work.

Big ups on your use of GGplot--best R graphing capabilities around! In response to your update about the QQ Plot, I didn't compare against normality like you did (the article is comparing exponentials, so a normal QQ isn't the best choice). The QQ Plot just compars the quantiles of one distribution to another (could be an ecdf against a hypothesized cdf, or an ecdf against another ecdf...). Essentially, by plotting one set of points against another, I'm suggesting that the empirical distribution of Annie is the same as the empirical distribution of Brian, or any other pairing.

Interesting paradox. I haven't seen that many statisticians using just a histogram when determining whether a certain distribution fits data reasonably. Kernel Density Estimators are a much better choice (for continuous data, like the data in the post), but they are also affected by your choice of bandwidth. When it comes down to it, like going to the doctor, sometimes the best choice is to get a second (or third!) opinion. For what it's worth, drawing a QQ Plot (something I've seen in every statistical consultation I've ever done) reveals the dependent structure of the data immediately and obviously in the form of a perfect linear relationship between any two variables.

There was always something very accessible about SCUMM games (not just LucasArts ones, but they certainly had some great ones). The story lines were simple and playful enough for kids to enjoy--even 8 year olds can understand the humor in an evil purple tentacle trying to take over the world, or a talking dog and rabbit solving crimes, or a quiet biker trying to repair his motorcycle.

It didn't take an intricate story or fast-paced action to set the LucasArts games apart from the competition. It was how simple, yet outlandish the plots were. 15 years later, I can still remember using a crowbar to remove a coin from the ground that was stuck there by chewing gum in Day of the Tenacle, or Ben Franklin naming a sandwich after a main character (Hoagie). I remember SCUMM games much in the same way I remember the plots of my favorite books I read as a kid. For me, they were much more like interactive novels or comic books than they were like video games. I remember Full Throttle, Dig, Loom, Monkey Island, and Day of the Tentacle in a very different way than I remember the other video games I played at the same age (Final Fantasy, Link to the Past, Super Mario World).

If you haven't had a chance to play any of these wonderful titles, I would highly recomend you grab the ScummVM and play some of the freeware titles (even if they aren't LucasArts): http://wiki.scummvm.org/index.php/Where_to_get_the_games.

Are there any plans to put this package on CRAN?

Sample size is something that must be interpreted in the presence of power. You can make a solid conclusion with a very small sample size if the true difference in arrival times is very large, given that the assumptions of the hypothesis test hold (t-test can be a little ridiculous with some of its assumptions sometimes).

In the original article, one of the footnotes mentioned that they tested the data using Wilcoxon's Signed-Rank test, which mitigates a lot of the impact of single outliers.

I'd love to see the raw data though, to see an even less-sensitive method to outliers (sign test). If the difference between the groups is as large as the article would lead us to believe, the loss of power should not present any problem.

You're going to open a big can of worms with that one ;). Tmux has quite a few more features than Screen, with Screen having a few features that are useful in edge cases[1]. It's hard to argue with tmux having bindings for vi/Emacs, as well.

If you're on a mac and use iTerm2, using tmux is somewhat of a no-brainer, thanks to the integration between tmux and iTerm2[2].

1: http://sourceforge.net/p/tmux/tmux-code/ci/master/tree/FAQ 2: https://code.google.com/p/iterm2/wiki/TmuxIntegration

The name choice seems bizarre to me. It's like a portmanteau of Fedora and Ubuntu, yet has very few of the features of Ubuntu. Judging by the default GUI, Facintosh seems more appropriate.

Gelman really is an unbelievable Bayesian. I don't think I'd say the most important since Bayes himself (Jeffreys and Jaynes strike me as crucial to the Bayes movement, along with the computational folks, like Geman&Geman and Metropolis&Hastings), but certainly one of the most important living Bayesians.

Gelman's (and Rubin, Carlin, Stern) Bayesian Data Analysis is by and large the most useful Bayes book out there. I keep it around just for reference sometimes, because of how detailed it is.

The argument seems very thin here. How is the mathematical equivalent of "blinding me with science" any proof that mathematics is not a language? I'm not arguing one way or another, but this does not demonstrate the proposition made in TFA.

This is the main reason I use git on my personal projects. I've found that it's quite difficult to pitch this feature to my group. The cost of learning an entirely new tech solution outweighs the benefits of version control in their minds--a statement I disagree. I have no killer feature for them in SCM that will make them want to start using it ASAP. Whether email is "good enough" has been made to be a matter of opinion, unfortunately.

Yes, there is. Link: http://domino.research.ibm.com/comm/wwwr_thinkresearch.nsf/p...

Not entirely related to Google's idea, but not completely unrelated either.

It's also much easier to release updates when all devices that run your OS are consistent, or at least largely homogeneous. Android smartphones have so much variety, both in terms of features and technical specifications, that it must be difficult to have updates released so quickly after Google pushes out a new version of the AOSP. AndroidCentral had a good article about this several months ago (link: http://www.androidcentral.com/why-you-ll-never-have-latest-v...)

I know one will argue it's not a viable solution for the 99% of Android phone owners out there, but the modding community for Android is very quick to update most flagship phones on most carriers, and even develops on phones that carriers have stopped pushing updates for. My Galaxy S1 had Ice Cream Sandwich on it when I finally traded up to a Nexus 4. It's not as convenient as Apple releasing the official update and rooting+customrecovery+clockworkmod+customrom is a process complicated enough to bar many people from entry, but for those capable folks, we can get updates almost as quickly as those on the Google flagship phone-of-the-day.

ESL is pretty much the go-to text for statisticians interested in machine learning. The math in it might be a bit more difficult than some of the pages you see on HN, but you won't get much better than this for a survey-course on machine learning from a statistician's perspective.

"To be fair, they do have windows in their office." Nice wordplay.

Not all work we do can be enjoyable, unfortunately. I bet this is a fun jab to bring up with him at parties. "Remember that time Wells made you make that security gif?"

Hah! Tim Pope is a pretty funny guy. His Vim plugins are fantastic as well (Pathogen and Surround are indispensable).

I work at a Fortune50, and I have to admit that if Mayer manages to turn Yahoo around that this will set a dangerous precedent for other large companies.

The ability to work remotely is a huge perk to having a tech job, and if other corporations see Yahoo killing WFH and then managing to save such a large sinking ship, they might infer causality.

Until Yahoo actually becomes successful as a whole again, this is just them thrashing about desperately. I'll be worried if we start to see success in Mayer's methods.

Thanks for the info. I'll have to check this out.

The interesting bit about this to me is not the interface, but the fact that they got it running on Android. I was curious in seeing if R could be ported to Android a few years back, and thought I remember reading there were some roadblocks pertaining to getting a decent BLAS to work on Android and iOS (which doesn't have a fortran compiler, iirc). The fact that Octave runs on Android means they must've figured out how to get a BLAS running on Android. Does anyone know what it's running? NetLib?

I always got the feeling people used Octave as "the next best thing" to having Matlab. Not everyone can shell out the money to MathWorks, especially outside academia, without having an employer pay for it. I used Matlab in college and found Octave to be mediocre. The speed leaves a great deal to be desired, especially if you've grown accustomed to the Parallel Toolbox and the inherent parallelization of the Statistics Toolbox since Matlab 2011a.

As for Python/R/Julia: Python is more general purpose than Matlab/Octave, though I have seen a push for NumPy and SciPy. R and Matlab work in somewhat different fields (exceptions to the rule: Functional data people in statistics seem to like Matlab). I always got the feeling the math/linear algebra/engineering people went with Matlab, while us statisticians went with R. There's plenty of information on Google about why that divide exists. Julia is more direct comparison to Matlab, and the only reason I can offer why people would not use it versus Matlab is age and wealth of libraries. Julia's a very new language, relative to Matlab, so we may see a wider adoption in the coming years from the Julia project.