Joe Blaylock

In various parts of the world, today is Veteran's Day, Armistice Day, or Remembrance Day. Here are some thoughts before I go to sleep.


Dear Veterans,

Thank you, sincerely, for your willingness to sacrifice everything for a bunch of spoiled slobs like me. Your resolve to protect what you love is woven deep into what it means to be human. The rest of us probably don't deserve you, and you've probably not gotten what you deserve.

Dear Soldiers,

I make it no secret that I'm a pacifist. Seriously. I wish you were out of a job. But until the day comes when the world can get along without heroic sacrifice, I remain awestruck when I think about the sacrifices you've chosen to make. Wherever you are, people who love you are waiting for you. Please be safe.




There's a strange tension between growing up in a family with rich ties to the US armed forces and the Buddhist call to compassionate action as I understand it. I fear that the soldiers in my life may not understand that my criticism of violence isn't a criticism of them, but of a global geopolitical system. That my own choices don't imply judgment of their choices. And I fear that my spiritual brothers and sisters of the Buddhist sangha will think me hypocritical. If they do, though, it's simply because they misunderstand how I can be saddened by the work at the same time that I admire the workers.

So apparently, you can attach certain classes of chemo drugs to peptide chains, and you can make them water soluble. This means they can get into tumors easier. So you get better poison delivery and less drug needed.

http://futurity.org/health-medicine/nano-scale-delivery-may-offer-less-toxic-chemo/#more-5366

I wonder if it will pan out. There seems to be a three-sided battle going on between people who want radically different ways to treat cancer - like delivering gold particles to tumors and then microwaving them, people who want to improve conventional chemo therapies, and people who want to find some kind of overall cure for cancer. The cure folks seem to be losing out to the other camps - as long as the widely looked-for "increasingly early diagnosis" shows up on time.

I'm excited by all of this, but disappointed it isn't happening faster.

So at the behest of SLAC Medical, I went to the allergist yesterday. I got the full monty: skin tests, asthma screening, training videos, and set up an appointment to consult about allergy shots.

Perhaps unsurprisingly, my allergies have continued to change as I've gotten older. My mold test was negative, and so were feathers. And the dust mite reaction was so mild the nurse wasn't sure whether she should count it as a hit or not. On the other hand, my reactions to weed pollens and grass continue to be off the charts. And to eggs.

We'll talk about the shots in a few weeks.

This weekend, moconnor, bethmolnar, anami and I went (via various circuitous routes) to Moab, Utah, for "The Other Half". Matthew and Beth had trained, and it looks to me like their first race went great. Anami, of course, had to take it off because of her busted ankle. I ran despite not having trained, and I did finish, so that's something. But there was (and is) some pain involved. Oops.

It was a pretty race, though. It was pretty much two hours and forty-five minutes of this:


birkwelch helped me out a lot. He had brought an extra iPod nano with him, and loaned it to me. His running mix is a masterpiece of excitation management. Notwithstanding the J-pop.

It's been fantastic hanging out with Sally and Anthony and Shanti and Matthew and Beth, getting to hang out and talk and eat nice food.


Oh yeah, and by the way:
Happy birthday Sally!

My birthday is Wednesday, and I'd like to have folks over here at our home in Redwood City next Saturday evening from around 7 until around whenever. I'm thinking drinks, cake, and conversation. Who's in?

Stanford researchers think they have a solid-state technology for detecting a variety of diseases (particularly cancers) that has about three orders of magnitude (that's 1000x) more sensitivity than current methods.
http://news.stanford.edu/news/2009/october12/cancer-detection-101209.html

I don't think I've been this excited about current research since http://bit.ly/ZgceM was still news. Of course, Faustman's spleen regeneration protocol hasn't really turned into a diabetes cure yet, and her initial results are almost ten years old at this point. Sometimes new techniques that look good just don't pan out. And sometimes it takes a long time to get things that do pan out into the hands of the public. And, sadly, sometimes research organizations' PR flaks exaggerate.

Still, I'm optimistic. Magnetic field detection and manipulation isn't exactly new science, and it would certainly dovetail well with cooking cancer without chemo. All the buzz seems to point to a consensus that we'll never cure cancer - but that we might be able to detect it and kill it with such speed and accuracy that it becomes a manageable nuisance.

This kind of technology could lay the groundwork for the Star Trek future where you spit in a cup, go get a cup of coffee, and when you come back, they tell you you've got about a dozen cancer cells, and here, drink this and then stand over in that funny shower looking thing. Spit in the cup, go drink more coffee. Come back, and they pronounce you cured. I think if it ever happens, it's many years away - but I really think it could happen. And the thought makes me feel like dancing.

Yay science!

So, as you may have noticed, anami and I are planning to go to Spain in December. We've got tickets into Barcelona, and then at the end of the trip, tickets to Cincinnati for HappyThanksNewChristYearMasGiving at sis's house. Anami and I want to climb while we're in Spain, and so we've got a fair bit of prep to do. The thing looming largest on my horizon at the moment is that I have to learn, at mininum, to set anchors for top rope.

Part one of setting anchors it just having someone show me how to do it. This shouldn't be a problem; I'm signed up for an REI class on the 24th, and if everything goes smoothly, I should pick up the basics there.

Part two, however, is trickier: I need to practice. Between now and December 16, I'd like to set as many bomber anchors as possible, and get folks to double-check me, critique my technique, and then climb on them. This is where you (the collective, impersonal you, which may or may not include the entity reading this) come in. I'd like to get outside, and I'd like to get outside with folks that are in a position to help with any of the following:

• re-teach things that REI taught me wrong


• teach me new things REI neglected


• places to go, and help getting there</a>
  
• folk that want to climb outside bad enough to actually show up


• belays


• maybe a small amount of healthy competition

Now, the way I figure it, I've got seven weekends between the 24th and time of departure. In that time, I want to get outside as much as possible, practice anchors in a variety of conditions (bolted and unbolted, taking advantage of a variety of terrain, on different types of rock) and, oh yeah, pick up any gear I still need (which is at least a stack of quick draws and some slings, but probably includes more stuff besides). I'm thinking nearby is good, to make it easy: Castle Rock, Pinnacles, Mickey's Beach? I dunno. I haven't been to any of these places yet.

So if you're interested in helping with this project - or you just want to get out on someone else's toprope, and you trust my safetiness - please let's do something soon. Stay in touch.

I baked my first by-myself bread tonight. Two loaves: one cinnamon raisin, one sundried tomato basil. Both are tasty, though they didn't rise like I expected. And Anami prefers something less doughy - I like a nice soft doughy loaf though. So this may be a point of contention. :)

In other news, Anami sprained her ankle, and needs to stay off her feet for a while. Anyone in the Bay Area that can drop by our place to help her out while I'm gone would be greatly appreciated.

Flying to Geneva on Wednesday. I'm nervous about the length of the flight. I'd prefer if I could be in the air for, say seven hours and then take a train for seven, or something. Anami assures me I'm being stupid. I agree. But that doesn't stop the trepidation.

On the other hand - Switzerland!

"What's particle physics?" asked Bod.

Scarlett shrugged. "Well," she said. "There's atoms, which is things that is too small to see, that's what we're all made of. And there's things that's smaller than atoms, and that's particle physics."

Bod nodded and decided that Scarlett's father was probably interested in imaginary things.

--From The Graveyard Book, by Neil Gaiman

The first time I came across an American Girl catalogue, I was a little bit whigged out. Even their black doll are clearly of mixed anglo descent - all of them have the exact same facial features, regardless of coloring. And where's the Chinese laundry worker of the late 19th century? Or the textile working American Girl, who makes all the pretty clothes for all the other American Girls, at the cost of some of her fingers?

But then I took some time to think it over, and I realized what the real goal of the line must be: to encourage young people to embrace period costume.

And then I decided that they were probably ok. Because really, the world needs more period costume.

Dear Lazyweb,

Back in Bloomington, there was this thing called the Bloomington Bike Project. It's a nonprofit that teaches people bike repair skills and offers those skills to the general public. One of the neat things about it is that after you volunteer a while, you can generally earn a bike. Kind of like Habitat for Humanity, with transportation.

anami is in Salt Lake City this week visiting family and running a marathon, and she stopped by the The Salt Lake City Bicycle Collective, where it's possible through some means not entirely obvious to me from looking at the website, to learn all the skills necessary to build a custom bike. And then to use those skills with their trash bike pile to build yourself a bike.

These things sound like fun to us, and probably way more practically useful than taking a bike repair class at REI or something.

So finally, the question: Is there something that functions similarly to these in the area around us? We're in Redwood City, and something nearby would be ideal. But going as far as Berkeley once a week probably wouldn't be completely out of the question.

What we're looking for is probably an environment to volunteer in, to learn bicycle repair by doing it, with the goal of learning enough to assemble custom bikes and pay only for the (very cheap because donated) parts used. Or a place where we can pay a small fee to learn a lot more than REI will teach us, with an eye towards building some bikes. Or similar variations - use your imagination and make suggestions.

Come on, lazyweb. You've been good to me before; don't fail me now.

Your pal,
Joe

Dear Lazyweb,

Somebody in the world must have done rigorous quantitative analysis comparing the various airline mileage plans and their associated credit card tie-ins to help you judge which one gets you free flights the fastest. I realize that this kind of research is precisely why the airlines make their websites so hard to navigate, but c'mon. Throw me a bone here. One of you must have seen such a thing.

If not, then tell me what airmiles program you're part of, and what you think of it. Bonus good will for suggestions about how I might game the system.

Thanks!

Your pal,
Joe

PS - comparing how fast you get points in points programs is nice and all, but at the end of the day, what really matters is dollars spent to get a free airfare - that is, ideally it should take into consideration what the points are actually worth (relative to other programs).

PPS - Somebody could use WWW::Mechanize and friends and do a startup around this, and I would totally browse their ad-supported website. It would rock.

Look what anami got me to celebrate my new gig!



http://farm3.static.flickr.com/2575/3828283088_e157715a06.jpg

I think I should take up golf. Or maybe sailing.

I got a job.

I signed an offer with the SLAC National Accelerator Laboratory on Tuesday. I'll be working on INSPIRE, the next-generation replacement for the current de facto document repository of the high energy physics community, SPIRES. Lots of computer programming and chillin' with librarians and physicists and stuff.

But wait! There's more!

INSPIRE is actually a set of extensions built on top of CERN Invenio. Which means, ladies and gentleman, that I'm now being paid to work on Free (as in freedom) Software. This pleases me greatly.

Ever since I got furloughed, money's been a little tighter than it used to be. Consequently, anami has been cooking a whole lot, to keep us from going out, and to help us economize with ingredient re-use. We also started shopping at the discount market here in Redwood City (called "Discount Market"), which is great. Today, she got 3 packages of Boursin (my favorite cheese, probably) for $3US. For those of you playing along at home, this is a steal.

And tonight, she cooked with it. "Well Dressed Shrooms" is mushrooms, gnocchi and onion in boursin, with fresh tomato and basil.

I ate rather more than an intended serving, and afterward lay on the couch gurgling with happiness.

Following up on a conversation with Luke Closs, I started thinking about measuring Wiki health out in the part of the world that doesn't keep track of RecentChanges. He pointed out that even in the RecentChanges world, as Wiki use becomes heavier and people integrate them more into their routines, RecentChanges use tends to drop off. People who are actually viewing and editing things just view and edit things - they know which parts have changed recently, because they're paying attention. Luke suggested that some combination of Views and Edits might be just as predictive of Wiki classification as RecentChanges - or perhaps useful in a different way.

So on Tuesday I sat down to figure it out.

I modified my previous work such that the k-means clustering takes place both with, and without the RecentChanges column in the data. I ran the clusters 1000 trials with 150 iterations for settling in each trial. By the time this is posted, the code should be in gitHub. The results for the smallest cluster in each trial are compiled into sums, and the sums saved to output files. This gives me the identifier for every wiki that has appeared in the smallest cluster of a given trial, and how many times they've been included. Intuitively, this should yield the set of the most similarly classified wikis over time, and if the numbers are similar between the two conditions (both with, and without, RecentChanges), then we can expect that RecentChanges, while capturing variance well, isn't necessary to predict wiki classification.

And so it is. The runs over the two conditions were correlated at a rate of 0.99; I consider it reasonable to work on wiki classification without recourse to RecentChanges.

This has the following immediately obvious results:

1. We can project the two remaining principal components into a 2D display space, so we can make nice scatter plots of the data and see if we can see the clusters ourselves.


2. All this stuff I'm doing is germane to the wider Wiki world where RecentChanges aren't available.

There are probably other effects as well, but I haven't seen them shake out yet.

In playing around with this, I generated some additional questions. I'm hoping that conversations with wiki pros can clear some of them up:

• I would expect Views to be artificially inflated, because users would go to a page, see it, *then* edit it. Should it be safe to assume that Edits can be factored out of Views? For example that we can have "Adjusted Views" = Views - Edits? And if my intuition that we can is reasonable, what does it mean when something has an Adjusted Views that's negative? Spambot?


• If it indicates a spambot, I think I've discovered a cheap way to detect spambotting.


• Do the unsupervised clusters actually reflect reality? Can I get some set of experts to provide classifications for, say, a hundred wikis, and then we can run a supervised method and compare these results?

We've all heard of the Anthropic Principal. You know, the one that asserts that the universe has to have the shape it does, because otherwise it couldn't support the existence of beings like us, who could notice its shape? Another way of looking at it is to say that there exist an infinity of possible universes, and we exist in this one because we can't exist in any of the others.

Today I'd like to propose that the anthropic principal is way off base, though. It ignores the Four Noble Truths, which I take to be something as strong as physical law. So a physics model that bottoms out at the anthropic principal is a) contrary to Buddhist theory, and b) contrary to human observation. But wait! It's easy to fix!

I propose the misanthropic principal: The universe has the shape it does because otherwise, there would be no suffering.

Shortly after getting the principal component analysis of my sample data set working, I also got k-means clustering working. As far as I can tell, the cluster quality looks pretty good.

Sample code is up on GitHub: jrbl/wikilytics.

To use it, you'll want a dump of the descriptive statistics for two or more wikis in a CSV format. For the 2098 wikis in my sample data, my dataset includes the following statistics:
* Wiki UUID
* RecentChanges count
* Page View count
* Page Edit count
* ViewedPages (count of distinct pages)
* Viewers
* Edited Pages (count of subset of pages receiving the edits)
* Editors (number of distinct editors, out of registered users)
* CommentedPages (number of pages with comments)
* Feed
* New
* Search
* Attachments
* Email
* Export
* Result of dividing ViewedPages/EditedPages (or 0)
* Result of dividing Viewers/Editors (or 0)

Clustering to two centroids gives a large set of wikis that are clearly unhealthy (every stat is at, or very close to, 0), and a slightly smaller group of wikis that are harder to classify. Three centroids still gives largely ambiguous results too. But when you cluster to four centroids and take the smallest set, you get back a list of wikis that looks good to me. When I eyeball their values in the spreadsheet, the values for the 3 principal columns map to my expectations about health.

The four clusters (on my most recent run) break down into groups of sizes: 1290, 16, 165, 627

Examining that group of 16 closer, we see the following values:

1. RecentChanges, Views, Edits
   
2. 46,1856,268
   
3. 211,1557,380
   
4. 0,133287,19
   
5. 7,1199,1
   
6. 43,4388,783
   
7. 27,4727,62
   
8. 27,23987,144
   
9. 81,19220,405
   
10. 919,3410,231
    
11. 3,2281,31
    
12. 25,7207,29
    
13. 81,3039,687
    
14. 6,2688,3
    
15. 232,1520,327
    
16. 1013,9530,563
    
17. 15,26550,7837
    

As we can see, all of these items have nonzero RecentChanges, except for one, which has a very large (comparatively speaking) number of Views. From what I know about collaboration and knowledge sharing tools, I would expect this set to be the wikis that we can call "Healthy". Of the other clusters, the size 1290 one is obviously the "Long Tail" of Wikis without users or traffic - most, or all of the data fields for most or all of these are zeros.

The other two groups are more interesting. They occupy a shadowy realm of wikis-which-might-be-healthy. Or perhaps wikis-for-very-particular-target-groups. I'm hoping to get a chance to meet with some luminaries of the social software space in the coming weeks and talk over whether there is - or can be - any good single characterization for these collections of items.

Besides that, the obvious next steps (insofar as purely numerical methods are concerned) seem to involve refining the k-means clustering and experiments on the one hand, and trying for a reliable, deterministic clustering, probably via some kind of simple SVM.

In both cases, my thinking is that the probabilistic nature of k-means is a little troubling. One approach to make the effect less onerous might be to cluster the data iteratively, doing a best-of-n for which wikis end up in each cluster size rankings. Another approach might be to calculate the variance predicted by each principal component of the data, keeping enough components to predict, say, 95% of the data variance, and throwing away the rest. In big-oh notation, this doesn't save you a lot of work for a pairwise Euclidean distance estimation, but it does save you some. You could also just trim that long tail before doing the classification - it's easy to tell what things go in that set without having a fancy algorithm. That about halves the number of data points to be clustered, which in an exponential algorithm is a big big win.

Lest I forget to mention it publicly, Eugene pointed out that Luke Closs initially suggested that RecentChanges might be pretty important. Looks like he was right. Thanks, Luke.