Information visualization is extremely important to the three major sports (American football, baseball, basketball) in the United States, as a sizable portion of each team's revenue derives from television contracts. The ability of the major networks to deliver information to their viewers at home is crucial to maintaining loyal viewers, and attracting new ones. Visualization of statistics is arguably most important to American football, hereafter referred to as football. With the plethora of rules, frequent rule changes between seasons, wide discrepencies between relevant statistical categories for each position, and relatively short careers of many players, football broadcasting needs easily digestible charts and graphs to make sense of the abundance of information. However, I've grown tired of the same standard bar graphs, tables, animations, and light-pen diagrams. These visualizations are limited in that they only compare one or two stats between one or two teams. I want visualize more stats about many teams, at the same time, and show relationships or correlations between several categories at once.
Thus, the goals of this project are two-fold: First, to explore various Parallel Coordinate tools as candidates for encoding more information about football statistics, and second, to test the validity of certain long-held notions about the game.
My Dataset is NFL football data for the 2000 season that ended in January, 2001. I gathered the data from ESPN.com and from NFL.com . Statistics for previous seasons are not readily available in digital form, and often are not available free-of-charge. This seems to be because gamblers and fantasy football enthusiasts will pay quite a lot of money for this type of information. At any rate, I have 19 statistical categories for 31 teams. My main goal is to establish Parallel Coordinates as an effective and useful method of visualizing this kind of information. This is a small data set, and I would need to conduct a more thorough analysis using more data to draw any general conclusions
-
Following is a typical bar graph one might see during
halftime of the broadcast of a football game. This graph
represent the total yardage earned on offense for each
team with 11 or more wins this season. The labels at the
bottom are not correct, as I created this with SpotFire
and didn't know how to manipulate the features of the bar
graph package. The color of each graph is encoding
another unimportant stat, but I basically used the
coloring to mimic the technique of making each team's bar
graph match their team colors.
Now, with Parallel Coordinates, I've encoded the same information about total yards of offense and wins. I have also encoded additional information about other offensive categories, including a breakdown of rushing offense and passing offense, as well as total points scored.
For this visualization, only the 12 teams with 10 or more wins are visible. We have eliminated the other 19 teams using the slider bar for the wins category.
- Many announcers, sports writers, and commentators make
the claim that the teams that win tend to be the ones that
can run the ball on offense, and defend against the run on
defense. I want to verify this visually.
We can see a stronger positive correlation between run defense, run offense, and wins than between pass defense, pass offense, and wins. This sample is too small to draw any general conclusions, but with more data we could try to visualize and verify this claim in general.
For this visualization, I left in all of the teams, but highlighted the teams with 10 or more wins. This clutters the display somewhat, but places the better teams in context with their peers.
- "Offense wins games, Defense wins championships."
We've heard this before; now let's show it scientifically,
visually, based on data and statistics!
Based on this visualization, there does not seem to be a stronger correlation between defensive rank, defensive yards per game, and wins than with the offensive categories. If we choose only the truly elite teams, i.e. teams with 11 or more wins, we see a stronger correlation between defense and winning. However, this particular visualization is important because it seems to show an unexpected inverse relationship between defensive and offensive rank, and defensive and offensive yards per game.
If I had more data on player salaries and time of possesion, I would like to delve into this subject more carefully. Announcers often claim that the defense gets tired if they're on the field too much, while they never make this claim about the offense. It's possible that there is a statistical basis for this seemingly irrational statement. Furthermore, I will bet that teams with a better defense probably have higher player salaries for the defensive squad than for their offensive unit.
-
Can we easily and visually compare statistics for the
entire league? Here is a link to all 19 categories for
all 31 teams, overlaid with Parallel Coordinates. Here is a screenshot of the
slider bars (simulating a dynamic query) which gives us
some idea where the top 8 teams lie in relation to the
rest of the league.
It is very difficult to discern correlations with this much data. With parallel coordinates, there are too many lines to discern patterns or correlations, and with the sliders we have no notion of how dense each range is.
The trellis function lets us plot all the data without overlaying. If we pare down our search space to only the offensive categories, we can easily visualize and compare the entire league.
We can also do the same for the defense. Note that in this visualization, the 12 teams with 10 or more wins are all highlighted.
Finally, here is a trellis plotting only three things: wins, running defense, and running offense, in an attempt to discern which is most strongly related to winning. You can tell by the "slope" of the line and its relative height: if it's sloping up from right to left, this suggests that rushing defense is more important rushing offense. Whereas if there is a dip in the middle, this suggests that rushing offense is more heavily weighted towards winning. If the line is flat, this suggests that a balance is effective.
Note that the trellis function isn't necessarily supposed to be used for this. However, I think it is an excellent technique, and should be expanded.
- Easy to import data from an Excel spreadsheet.
SpotFire can directly open an Excel spreadsheet, and it
"does the right thing" with the data.
- One requirement for this visualization is that rankings
and defensive data need to be "inverted", i.e. being ranked
number 1 is better than number 31, and allowing only 200
points on defense is better than yielding 800 points.
Spotfire handles this easily, and can do so for multiple
axes at once.
- It is very difficult, if not impossible, to change the
data type of a category. Even after explicitly setting the
cells of the Excel spreadsheet to integers, Spotfire
imported all of these values as doubles. There is no need
for double-precision floating point numbers on a range
slider tracking the number of wins between 1 and 13.
- You cannot force Spotfire to display labels on every
parallel axis. There are options you can set that tell
Spotfire to do so, but they are not always enforced. You
can recover the label information with a mouse-over, but you
cannot export visualizations with all X-axes labeled. These
visualizations are flawed, because an exported image loses
the mouse-over feature!
- You cannot reorder the columns by drag-N-drop in
Spotfire. This would be extremely useful to quickly check
for several correlations.
- You cannot adjust the granularity of the slider
bars. Some categories could slide in units of 100 or 1000,
but there is no mechanism for setting this.
- It's not clear how to un-select things. I figured it out, but it took me a few minutes.
- By highlighting multiple columns, you can select
all the elements that are in the range in any
of the columns. This is a logical OR operation. Its
counterpart, a logical AND, i.e. select only the
elements that are in the range for all the columns,
is not present.
- In sports, we often want to know if our team is the
"best of all time". This is a tricky comparison, as changes
to the rules between seasons make comparisons of raw numbers
accross the years suspect or inaccurate. One way of
measuring "the best of all time" is to measure how many
standard deviations above the league average a team is for a
certain category. It would be nice if Spotfire could insert
a bright red line that represents the mean for each
category. I could easily calculate this in an excel
spreadsheet, but then I would need to find it, color it
manually, etc. This would be a neat feature.
Here is a visualization of the entire league, with the league average manually selected.
For more information about standard deviations and sports statistics, read Eddie Epstein's insightful article, The Best Defense of All Time?
- I would like to be able to remove entire rows. Because
Parallel Coordinates does everything based on columns of
data, it is relatively easy to move and adjust columns. I
would like to be able to adjust rows, especially with the
trellis function, to directly compare a few data points
without overlaying them. You cannot, for example, select
only 6 data points and show then with the trellis function.
A link to my impressions of this tool.
Conclusions about visualizing this data with Spotfire.:
-
This is a fun project. Spotfire is a very richly featured tool.
-
I think you can use parallel coordinates for these types of visualizations, but it would require more data and more time to make any general claims.
-
I already knew what I was looking for. Discerning arbitrary patterns and unknown relationships in a fresh dataset seems difficult.
A Parallel Coordinates Applet by Amit Goel from Virgina Tech. It's written in Java, it's free, and it was home-grown by a grad student. Thus, it is not fair to pick on this tool. It does have one excellent feature which spotfire lacks.
The applet is flaky, tempermental, and effectively not extendable. Though it is written in Java, it uses deprecated AWT components from jdk 1.1.X and 1.0.X. The source is available, but does not compile under 1.1.8. The code seems to contain errors in addition to the deprecation warnings. Luckily, a jar file is available which works locally. Data must be imported in STF (Simple Table Format). This is a pretty standard format for many columns of data. You cannot import Excel spreadsheets, and the Applet does nothing rather than return an error message if your data file contains errors.
I was not able to capture any screenshots because I was not able to get a large enough data set into .stf format, and successfully import this data into the Applet.
This applet does have one excellent feature which Spotfire lacks: You can drag-N-drop columns! This feature allows you to quickly move and manipulate columns to search for patterns.