Network visualizations of passes in ultimate
As I have not proceeded much in the analysis of Team USA ultimate passing networks during the weekend, I will instead show what kinds of visualizations of passes one could work out with network visualization tools. Most of the pictures are created using NetDraw (a complementary visualization package to UCINET social network analysis tool).
In this first picture, players are the dots and passes between them are shown with arrows. The data is from the final game of Team USA against Australia.
If there are passes in both directions between two players, there is arrowhead in both ends. The thicker the arrow, the more passes were thrown between these people. There are some passes (3 total) to a person called “n/a”. These are throwavays by players, but the intended receiver was not recorded. Other all the other turnovers (like dropped passes) are shown is this picture. The coloring of the players does not make much sense here, but they show which players are part of different k-cores (a social network analysis term). The players in the center of the graph are more central (more connected).
The next picture shows players as circles and squares (data from the same game as above). This might get a little confusing, but throwers are shown as circles and receivers as squares (you might notice that the arrows only point from circles to squares). Again the same players (Deaver and Ziperstein are in the center - both as receivers as wall as throwers). Again the coloring is for showing the k-cores. In short blues are more connected than blacks, which are in turn more connected than reds.
In both of the pictures the layout of the players on the canvas is based on spring embedding calculations automatically done by the visualization software. There are also other ways to position the players like multi-dimensional scaling. The MDS works pretty well with teams with more players, positioning players who could substitute each other close together. However as in World Games only 11 players were allowed in each team, the players are more connected and the MDS layout tends to clump the players all together (as they have pretty much the same connections to other players). In a team were there is separate offensive and defensive lineups, MDS works pretty well.
There are also tools to visualize dynamic social networks (which fits well in theory to passing networks…). I have prepared an animation of passes in WUGC final last year (ladies game), but I need to tweak that one a little before it is ready for publicity.
Another way to visualize the game from a little different angle is to place players as throwers and receivers in a x-y plane, where y-axis measures how much yardage the player has generated for his/her team and x-axis shows shows what kind of participation the player has had in how the possessions have ended (goals are positive, turnovers are negative). The following picture is from womens final game in WUGC 2004 (Finnish team). The lines depict the players, round end is the player as receiver and the other (green) end is the player as thrower. The x-position is calculated by giving different weights for different types of turnovers (drops, throwaways, blocks/interceptions, stalls…). The y-axis is calculated by first categorising the passes (hucks, shorts, swings, etc) and then adding the passes together using different weights for different passes (so the y-value is not an absolute yardage). As I am missing the throwtype data on most of the throws in World Games, I canot show this kind of picture for Team USA, sorry.
I am not saying this kind of graph is the ultimate truth of offensive value of a player, but maybe it stirs some discussion. For example (as Jim P has already mentioned), this kind of picture does not differentiate players who are advancing the disc with a few long throws from players who are avancing the dis with many short throws.
August 29th, 2005 at 8:54 pm
I think something interesting to do with this passing data would be to scale it by opportunity. For example, Namkung and Deaver did not throw many passes to one another, despite both having a lot of touches. But this is unsurprising when you realize that Deaver played mostly offensive points, and Namkung played mostly defensive points.
Similarly, players on defensive points often never see the disc, so looking at who they (didn’t) throw to on that point would be fairly meaningless. So even better than looking on a per-point basis would be looking on a per-posession basis.
So, for example, when Deaver and Namkung are both on the field, Namkung throws .3 passes per posession to Deaver. Or something like that.
August 29th, 2005 at 9:22 pm
Good point - especially that one should use possession counts, not only counting how many times players are on the field (at the same time). I have not yet used the line-up data at all. But it’s there, so I am getting there at some point (working on this only late at night makes the progress kind of slow…)
Ok, decided to quickly produce a graph for Furious George from the WUGC 2004 final against Condors. In this picture the offensive players are on the right and defensive players are on the left (the layout is again automatically generated using spring-embedding). Evan Wood is very central, as he was playing in bith lineups and participated in offensive plays quite a lot. Defensive players have less passes among them than offensive players, that’s why their circles are smaller (and of different color). One cannot get this clear spearation with Team USA, as there were not enough players for two “separate”lineups.
