THE BOOK--Playing The Percentages In Baseball

I was doing some research on clustering last week in search of a Master’s thesis topic and thought something like http://www4.ncsu.edu/~cdwessel/NSMC%20Talk.pdf
might be interesting to use on pitch f/x data.

Jimmy,

The clustering research in PITCHf/x is quite far ahead of the chapters from the MacKay book that you recommend there.  Some of that research has been publicly discussed, but not all of it.  I’ve personally tried many clustering techniques that are more advanced than those mentioned in that text, as have other researchers.  Some of the algorithms show some promise, but all of them are problematic.  Most of the problems have to do with not handling noisy data well.  There are clustering techniques that handle noisy data better, but even they have some trouble.

Ultimately, it seems that you trade accuracy for expediency with a clustering algorithm.  That works okay if you are just wanting to make some decent clusters and draw conclusions about large populations.  However, if you’re trying to measure variance of clusters and get anything meaningful from that, the performance of a clustering algorithm is going to need to be a lot better than anything I’ve seen so far.  (I’d be happy to correspond offline about some of the more promising clustering techniques I’ve looked at.)

Hi Mike,

The McKay book is intended as an introductory look, certainly not as a definitive guide to clustering techniques. It only covers the very basics of k-means, and doesn’t even touch some other well-known ones like QT or annealing. The literature on clustering algorithms out there (took a peek last night) actually seems pretty huge as you noted, but I really am not up to date on it at all. Interesting subject though!

At any rate, my main point isn’t so much the clustering algorithm itself, since that subject is still evolving and, like you mentioned, still has problems to deal with. What I was trying to get at more directly was the fitting of distributions to the clusters, to rigorously quantify the spread. In my opinion there are tons of areas in baseball analysis like this where a simple technique from statistical theory can be brought in to change the discussion from qualitative analysis to quantitative analysis. And that means a lot.

In the end, I think for my specific example (at least) the problem might be mitigated with the help of human assistants, since a lot of times it’s easier for a person to positively identify a pitch as a slider versus a cutter, or whatnot. Then the clustering algorithm is even kind of moot and one can proceed with the rest of it.

Mike, I’d be glad to maybe look at some of the techniques you are alluding to in my spare time, just to satisfy my curiosity. Shoot me an e-mail?

gogurt at gmail

Much thanks for your input!

I have always thought--but unfortunately haven’t had the time to get into it--that Model-Based clustering would be a good candidate for unsupervised pitch identification.  Similar to the excerpt above, it allows for ellipsoidal and varying variances across pitches, and automatically chooses the optimal number of clusters based on BIC.  This allows the modeling of, for example, a new pitcher for whom we’re not sure of his pitches.  In this case, there is no need to choose k, and less worry about misplacement of the centroids.

Millsy, model-based clustering is the technique that I have had the most success with.  It still needs a fair amount of babying.