Sorry to say, but I disagree with the premise of this article. You still need R to answer many of the questions the author of the article is proposing to answer.
On top of that, having a degree in AI will prepare you for a lot of real-world problems because you've been lots of data sets. I'd prefer to have an PhD-level researcher who knows why my support vector machine is overtrained, than a guy with real-world experience who is likely to overtrain and bias the model.
Furthermore, the author's example of weight and height correlation isn't one of a bad model. It would be a perfectly fine model if that was all the data you had. If you ignore lots of input data to build a crappy model, you're going to have a bad time... But everyone with a PhD and now "real world experience" would know this.
The title of the article is misleading. One may think that R is a tool is worse than other alternatives. It's more like "current methods (or simple methods) are not enough". R is as good for data science as any other tool (and arguably is even better thanks to its extensive library of packages).
I think Big Data is about convincing companies they need to upgrade their "tools" and purchase your software or services, or convincing managers in other departments you, the Big Data team, are doing valuable work. It's primarily hype-driven. Classic IT.
Most money will be made not by exploiting actually the "big data", but by exploiting the perceived need to collect and exploit it. We'll succeed in the former, but not necessarily in the later. It's more about keeping pace with one's competitors. "If they're doing it, we need to be doing it too."
Lots of meetings with presentations that will use "big data". But few results that ever come out of that data.
Amazingly bad article, riddled with inaccuracies. Is it because the stat/math has to be toned down so much becaudse otherwise the average Forbes reader can't grok the content ?
First he says, "mean and standard deviation define a normal distribution". They do no such thing. You don't even know what probability distribution attendance at sports events is when you compute mu and sigma. Then he says, ok its not normal, its power law. But the reason he attributes for power law is all wrong. Its not normal because there are no "negative people" to lower the mean ?! Ha ha ha!
His second example on the height weight regression line has little to do with whether there is a physical relationship between height and weight. If you really went to know, your weight might be a function of some two dozen variables like body surface area, bone structure/density, genetics, food-intake, income levels, nature of work ( sedentary vs active ) etc. Even if you gathered data on all those, 24 dimensions is a lot. So you do a principal components, and what does that do ? The first PC from the PCA gives you a conflated component that explains say 90% of your weight. That conflated component is going to be a linear weighted sum of many of the 24 dimensions. That conflated component is just a math object, doesn't exist in reality. If you want, I can easily connect your body weight to the attendance at your local sports event and a residual, and that regression might even be a good fit :)
In any case, none of this has anything to do with R, so his title is pure linkbait.
I was really hoping this would say Big Data people are moving towards Python -- or, i prayed Ruby. But nah, this article had nothing to do with R or the others.
8 comments
[ 2.8 ms ] story [ 25.0 ms ] threadOn top of that, having a degree in AI will prepare you for a lot of real-world problems because you've been lots of data sets. I'd prefer to have an PhD-level researcher who knows why my support vector machine is overtrained, than a guy with real-world experience who is likely to overtrain and bias the model.
Furthermore, the author's example of weight and height correlation isn't one of a bad model. It would be a perfectly fine model if that was all the data you had. If you ignore lots of input data to build a crappy model, you're going to have a bad time... But everyone with a PhD and now "real world experience" would know this.
twitch
http://cscs.umich.edu/~crshalizi/weblog/491.html
http://cscs.umich.edu/~crshalizi/weblog/390.html
Most money will be made not by exploiting actually the "big data", but by exploiting the perceived need to collect and exploit it. We'll succeed in the former, but not necessarily in the later. It's more about keeping pace with one's competitors. "If they're doing it, we need to be doing it too."
Lots of meetings with presentations that will use "big data". But few results that ever come out of that data.
His second example on the height weight regression line has little to do with whether there is a physical relationship between height and weight. If you really went to know, your weight might be a function of some two dozen variables like body surface area, bone structure/density, genetics, food-intake, income levels, nature of work ( sedentary vs active ) etc. Even if you gathered data on all those, 24 dimensions is a lot. So you do a principal components, and what does that do ? The first PC from the PCA gives you a conflated component that explains say 90% of your weight. That conflated component is going to be a linear weighted sum of many of the 24 dimensions. That conflated component is just a math object, doesn't exist in reality. If you want, I can easily connect your body weight to the attendance at your local sports event and a residual, and that regression might even be a good fit :)
In any case, none of this has anything to do with R, so his title is pure linkbait.