Comments for Win-Vector Blog http://www.win-vector.com/blog The Applied Theorist's Point of View Wed, 25 Jan 2012 05:27:08 +0000 hourly 1 http://wordpress.org/?v=3.3.1 Comment on Correlation and R-Squared by Ritesh http://www.win-vector.com/blog/2011/11/correlation-and-r-squared/comment-page-1/#comment-6048 Ritesh Wed, 25 Jan 2012 05:27:08 +0000 http://www.win-vector.com/blog/?p=1866#comment-6048 Great post. thanks for the succinct explanation. I always read that R2 is square of correlation but rarely anyone pointed out the details and I was never able to prove it to myself. Reading your post made things to clear me. Thanks Great post. thanks for the succinct explanation. I always read that R2 is square of correlation but rarely anyone pointed out the details and I was never able to prove it to myself. Reading your post made things to clear me. Thanks

]]> Comment on Six Fundamental Methods to Generate a Random Variable by jmount http://www.win-vector.com/blog/2012/01/six-fundamental-methods-to-generate-a-random-variable/comment-page-1/#comment-6037 jmount Sat, 21 Jan 2012 23:44:18 +0000 http://www.win-vector.com/blog/?p=1925#comment-6037 Getting some interesting feedback from good friends on this article. I wish I had more time to spend on: * Pseudo physical systems like low digits from stock prices. * Complexity and cryptography. * Bias removal (see: "Fast Simulation of New Coins From Old", Serban Nacu and Yuval Peres, The Annals of Applied Probability, 2005 vol. 15 (1A) pp. 93-115). * Information theory issues (how many bits you need to generate a structure, versus how long you need to run to be correct): "How to recycle random bits", Russell Impagliazzo and David Zuckerman, FOUNDATIONS OF COMPUTER SCIENCE, 1989 vol. 30). * Expander graphs and quasi random structures. Instead we are promising only a future write up on just one essential issue: how deterministic systems can even appear to be at all random (e.g. ergodic theory). Getting some interesting feedback from good friends on this article. I wish I had more time to spend on:

* Pseudo physical systems like low digits from stock prices.
* Complexity and cryptography.
* Bias removal (see: “Fast Simulation of New Coins From Old”, Serban Nacu and Yuval Peres, The Annals of Applied Probability, 2005 vol. 15 (1A) pp. 93-115).
* Information theory issues (how many bits you need to generate a structure, versus how long you need to run to be correct): “How to recycle random bits”, Russell Impagliazzo and David Zuckerman, FOUNDATIONS OF COMPUTER SCIENCE, 1989 vol. 30).
* Expander graphs and quasi random structures.

Instead we are promising only a future write up on just one essential issue: how deterministic systems can even appear to be at all random (e.g. ergodic theory).

]]> Comment on An Appreciation of Locality Sensitive Hashing by jmount http://www.win-vector.com/blog/2011/11/an-appreciation-of-locality-sensitive-hashing/comment-page-1/#comment-5632 jmount Fri, 23 Dec 2011 20:26:25 +0000 http://www.win-vector.com/blog/?p=1848#comment-5632 Some simple example code now up on GitHub: <a href="https://github.com/WinVector/Locality-Sensitive-Hashing-Example" rel="nofollow">https://github.com/WinVector/Locality-Sensitive-Hashing-Example</a>. It is just a simple example on random data- but even on as few as 10,000 vectors we see reliable results and a 25 times speedup over brute force. Also the simple code deals two major problems seen when working in Java: expense of many small objects and inability to reliably key things off floating point (due to non-repeatiablity default Java floating point). The detailed log given in the example show how we sweep all projection widths until we find a sweet spot where we are getting sets small enough to inspect (look at fraction inspected, when it is low we are skipping large components). Some simple example code now up on GitHub: https://github.com/WinVector/Locality-Sensitive-Hashing-Example.

It is just a simple example on random data- but even on as few as 10,000 vectors we see reliable results and a 25 times speedup over brute force. Also the simple code deals two major problems seen when working in Java: expense of many small objects and inability to reliably key things off floating point (due to non-repeatiablity default Java floating point).

The detailed log given in the example show how we sweep all projection widths until we find a sweet spot where we are getting sets small enough to inspect (look at fraction inspected, when it is low we are skipping large components).

]]> Comment on My Favorite Graphs by stefano ciotti http://www.win-vector.com/blog/2011/12/my-favorite-graphs/comment-page-1/#comment-5278 stefano ciotti Wed, 07 Dec 2011 09:51:08 +0000 http://www.win-vector.com/blog/?p=1886#comment-5278 It's a fantastic site. Well done!!! It’s a fantastic site. Well done!!!

]]> Comment on My Favorite Graphs by Naas http://www.win-vector.com/blog/2011/12/my-favorite-graphs/comment-page-1/#comment-5276 Naas Wed, 07 Dec 2011 09:00:51 +0000 http://www.win-vector.com/blog/?p=1886#comment-5276 Thanks for the post. It's very easy to understand and I'm sure that most predictive modeling practitioners will find it useful. I mostly use SAS but I'm very impressed by the flexibility and graphics capabilities of R. They must just work on the processing power (utilizing more than one processor) and the connectivity to RDB's. Both of which are good and easy in SAS but much more difficult in R Thanks for the post. It’s very easy to understand and I’m sure that most predictive modeling practitioners will find it useful. I mostly use SAS but I’m very impressed by the flexibility and graphics capabilities of R.
They must just work on the processing power (utilizing more than one processor) and the connectivity to RDB’s. Both of which are good and easy in SAS but much more difficult in R

]]> Comment on My Favorite Graphs by Nina Zumel http://www.win-vector.com/blog/2011/12/my-favorite-graphs/comment-page-1/#comment-5264 Nina Zumel Tue, 06 Dec 2011 18:04:21 +0000 http://www.win-vector.com/blog/?p=1886#comment-5264 Ah, very nice! Thanks, Ian. Here's the plot: <img src="http://www.win-vector.com/blog/wp-content/uploads/2011/12/Rplot1.png" /> Ah, very nice! Thanks, Ian.

Here’s the plot:

]]> Comment on My Favorite Graphs by Ian Fellows http://www.win-vector.com/blog/2011/12/my-favorite-graphs/comment-page-1/#comment-5255 Ian Fellows Tue, 06 Dec 2011 02:59:01 +0000 http://www.win-vector.com/blog/?p=1886#comment-5255 <a href="#comment-5254" rel="nofollow">@Ian Fellows </a> model.glm <- glm(formula=chd~ ldl*famhist,family=binomial(),data=saheart,na.action=na.omit) plot(effect(term="ldl:famhist",mod=model.glm,default.levels=20),rescale.axis=FALSE) @Ian Fellows

model.glm <- glm(formula=chd~ ldl*famhist,family=binomial(),data=saheart,na.action=na.omit)
plot(effect(term="ldl:famhist",mod=model.glm,default.levels=20),rescale.axis=FALSE)

]]> Comment on My Favorite Graphs by Ian Fellows http://www.win-vector.com/blog/2011/12/my-favorite-graphs/comment-page-1/#comment-5254 Ian Fellows Tue, 06 Dec 2011 02:57:59 +0000 http://www.win-vector.com/blog/?p=1886#comment-5254 I love to see the box and jitter plot getting some love. It is one of my GOTO plots. I also really like effect plots for logistic regression: library(effects) model.glm 0.0 ~ ldl*famhist,family=binomial(),data=saheart) plot(effect(term="ldl:famhist",mod=model.glm,default.levels=20),rescale.axis=FALSE) I love to see the box and jitter plot getting some love. It is one of my GOTO plots. I also really like effect plots for logistic regression:

library(effects)
model.glm 0.0 ~ ldl*famhist,family=binomial(),data=saheart)
plot(effect(term=”ldl:famhist”,mod=model.glm,default.levels=20),rescale.axis=FALSE)

]]> Comment on “The Mythical Man Month” is still a good read by jseliger http://www.win-vector.com/blog/2011/10/the-mythical-man-month-is-still-a-good-read/comment-page-1/#comment-5079 jseliger Sun, 23 Oct 2011 19:48:12 +0000 http://www.win-vector.com/blog/?p=1834#comment-5079 It's not only a good read for hackers, either: it's a good read for almost any group who have an intellectually demanding task that can't be easily parallelized. I talked about <em>The Mythical Man Month</em> in a blog post about grant writing: http://blog.seliger.com/2009/08/23/one-person-one-proposal-dont-split-grant-writing-tasks , because in the grant world nonprofits will try to divide writing tasks among a group. The result is a proposal that, even if it hits the deadline, is probably a bad proposal. I find books that cross their putative genre or field very interesting. Chris Matthews' <em>Hardball<em> is another; it's about politics, but it's really about life. It’s not only a good read for hackers, either: it’s a good read for almost any group who have an intellectually demanding task that can’t be easily parallelized. I talked about The Mythical Man Month in a blog post about grant writing: http://blog.seliger.com/2009/08/23/one-person-one-proposal-dont-split-grant-writing-tasks , because in the grant world nonprofits will try to divide writing tasks among a group. The result is a proposal that, even if it hits the deadline, is probably a bad proposal.

I find books that cross their putative genre or field very interesting. Chris Matthews’ Hardball is another; it’s about politics, but it’s really about life.

]]> Comment on Kernel Methods and Support Vector Machines de-Mystified by jmount http://www.win-vector.com/blog/2011/10/kernel-methods-and-support-vector-machines-de-mystified/comment-page-1/#comment-5073 jmount Thu, 20 Oct 2011 23:25:16 +0000 http://www.win-vector.com/blog/?p=1804#comment-5073 <a href="#comment-5071" rel="nofollow">@shoonya </a> Thanks! I used R and ggplot2 for all of the plots. For the SVM steps I just pasted the formulas from "Kernels Methods for Pattern Analysis" into http://www.win-vector.com/blog/2010/06/automatic-differentiation-with-scala/ and let a conjugate gradient optimizer do the work (not recommended for real sized problems). @shoonya
Thanks! I used R and ggplot2 for all of the plots. For the SVM steps I just pasted the formulas from “Kernels Methods for Pattern Analysis” into http://www.win-vector.com/blog/2010/06/automatic-differentiation-with-scala/ and let a conjugate gradient optimizer do the work (not recommended for real sized problems).

]]>