http://thecity2.com/2012/03/02/visualiz ... lustering/
Crow mentioned the Rockets in the other thread, so I'll use them as an example. They have a neat set of about 3 clusters:
The other 29 teams can be found at that link above.
Visualizing Team Clusters
Visualizing Team Clusters
Stuff at my blog on visualizing who plays with whom:
http://thecity2.com/2012/03/02/visualiz ... lustering/
Crow mentioned the Rockets in the other thread, so I'll use them as an example. They have a neat set of about 3 clusters:
The other 29 teams can be found at that link above.
http://thecity2.com/2012/03/02/visualiz ... lustering/
Crow mentioned the Rockets in the other thread, so I'll use them as an example. They have a neat set of about 3 clusters:
The other 29 teams can be found at that link above.
Re: Visualizing Team Clusters
Yeah the Rockets appears to have 3 complete squares. Some other teams have that, though most have just 2. A few just have 1 fairly to really intact one.
If one assigned players to player type clusters (using one method or another or a combination) and calculated RAPM player pairs (and trios, quads and 5 man RAPM) for the current season one could find league wide trends for player pair usage and find very large minute sample trends for player pair success and failure that might be helpful beyond what this data says about unique pairs with relatively small minute samples. I mentioned this basic thought to a certain GM (in Houston) almost 2 years ago. He wrote back the same or next day saying he thought that was an interesting idea. It was similar to some work David Sparks had done (that probably played a role in his getting involved with the Celtics) but my proposal was to use RAPM rather than a crude, incomplete boxscore metric.
If one assigned players to player type clusters (using one method or another or a combination) and calculated RAPM player pairs (and trios, quads and 5 man RAPM) for the current season one could find league wide trends for player pair usage and find very large minute sample trends for player pair success and failure that might be helpful beyond what this data says about unique pairs with relatively small minute samples. I mentioned this basic thought to a certain GM (in Houston) almost 2 years ago. He wrote back the same or next day saying he thought that was an interesting idea. It was similar to some work David Sparks had done (that probably played a role in his getting involved with the Celtics) but my proposal was to use RAPM rather than a crude, incomplete boxscore metric.
Re: Visualizing Team Clusters
The corrplot() package provides even more options and clustering methods. Check this out for GSW:
And here's one featuring last year's NBA Finals matchup between DAL and MIA:
And here's one featuring last year's NBA Finals matchup between DAL and MIA:
Re: Visualizing Team Clusters
I don't know how hard it would be but can you create a league composite, coding the players by position? I think it could be interesting to compare individual teams to the league composite to get a better sense of relationship strengths and weaknesses by position and location in the rotation.