Sports ranking methods, 3

This is the third of a series of expository posts on matrix-theoretic sports ranking methods. This post discusses the random walker ranking.

We follow the presentation in the paper by Govan and Meyer (Ranking National Football League teams using Google’s PageRank). The table of “score differentials” based on the table in a previous post is:

\begin{tabular}{c|cccccc} \verb+x\y+ & Army & Bucknell & Holy Cross & Lafayette & Lehigh & Navy \\ \hline Army & 0 & 0 & 1 & 0 & 0 & 0 \\ Bucknell & 2 & 0 & 0 & 2 & 3 & 0 \\ Holy Cross & 0 & 3 & 0 & 4 & 14 & 0 \\ Lafayette & 10 & 0 & 0 & 0 & 0 & 0 \\ Lehigh & 2 & 0 & 0 & 11 & 0 & 0 \\ Navy & 11 & 14 & 8 & 22 & 6 & 0 \\ \end{tabular}
This leads to the following matrix:

M_0=\left(\begin{array}{cccccc} 0 & 0 & 1 & 0 & 0 & 0 \\ 2 & 0 & 0 & 2 & 3 & 0 \\ 0 & 3 & 0 & 4 & 14 & 0 \\ 10 & 0 & 0 & 0 & 0 & 0 \\ 2 & 0 & 0 & 11 & 0 & 0 \\ 11 & 14 & 8 & 22 & 6 & 0 \\ \end{array}\right) .

The edge-weighted score-differential graph associated to M_0 (regarded as a weighted adjacency matrix) is in the figure below.

sm261_baseball-ranking-application_teams-digraph2

This matrix M_0 must be normalized to create a (row) stochastic matrix:

M = \left(\begin{array}{cccccc} 0 & 0 & 1 & 0 & 0 & 0 \\ {2}/{7} & 0 & 0 /{7} /{7} & 0 \\ 0 /{7} & 0 /{21} /{3} & 0 \\ 1 & 0 & 0 & 0 & 0 & 0 \\ {2}/{13} & 0 & 0 /{13} & 0 & 0 \\ {11}/{61} /{61} /{61} /{61} /{61} & 0 \\ \end{array}\right) .

Next, to insure it is irreducible, we replace M by A=(M+J)/2, where J is the 6\times 6 doubly stochastic matrix with every entry equal to 1/6:

A=\left(\begin{array}{cccccc} {1}/{12} /{12} /{12} /{12} /{12} /{12} \\ {19}/{84} /{12} /{12} /{84} /{84} /{12} \\ {1}/{12} /{84} /{12} /{28} /{12} /{12} \\ {7}/{12} /{12} /{12} /{12} /{12} /{12} \\ {25}/{156} /{12} /{12} /{156} /{12} /{12} \\ {127}/{732} /{732} /{732} /{732} /{732} /{12} \end{array}\right).

Let

{\bf v}_0 = \left( \frac{1}{6} , \frac{1}{6} , \frac{1}{6} , \frac{1}{6} , \frac{1}{6} , \frac{1}{6}\right).

The ranking determined by the random walker method is the reverse of the left eigenvector of A associated to the largest eigenvalue \lambda_{max}=1 (by reverse, I mean that the vector ranks the teams from worst-to-best, not from best-to-worst, as we have seen in previous ranking methods).
In other words, the vector

{\bf r}^*=\lim_{n\to \infty}{\bf v}_0A^n.

This is approximately

{\bf r}^* \cong \left(0.2237\dots ,\,0.1072\dots ,\,0.2006\dots ,\,0.2077\dots ,\,0.1772\dots ,\,0.0833\dots \right).

Its reverse gives the ranking:

Army < Lafayette < Bucknell < Lehigh < Holy Cross < Navy.

This gives a prediction failure rate of 13.3\%.

Sports ranking methods, 2

This is the second of a series of expository posts on matrix-theoretic sports ranking methods. This post discusses Keener’s method (see J.P. Keener, The Perron-Frobenius theorem and the ranking of football, SIAM Review 35 (1993)80-93 for details).

See the first post in the series for a discussion of the data we’re using to explain this method. We recall the table of results:

X\Y Army Bucknell Holy Cross Lafayette Lehigh Navy
Army x 14-16 14-13 14-24 10-12 8-19
Bucknell 16-14 x 27-30 18-16 23-20 10-22
Holy Cross 13-14 30-27 x 19-15 17-13 9-16
Lafayette 24-14 16-18 15-19 x 12-23 17-39
Lehigh 12-10 20-23 13-17 23-12 x 12-18
Navy 19-8 22-10 16-9 39-17 18-12 x
sm261_baseball-ranking-application_teams-digraph

Win-loss digraph of the Patriot league mens baseball from 2015

Suppose T teams play each other. Let A=(a_{ij})_{1\leq i,j\leq T} be a non-negative square matrix determined by the results of their games, called the preference matrix. In his 1993 paper, Keener defined the score of the ith team to be given by

s_i = \frac{1}{n_i}\sum_{j=1}^T a_{ij}r_j,

where n_i denotes the total number of games played by team i and {\bf r}=(r_1,r_2,\dots ,r_T) is the rating vector (where r_i\geq 0 denotes the rating of team i).

One possible preference matrix the matrix A of total scores obtained from the pre-tournament table below:

A = \left(\begin{array}{rrrrrr} 0 & 14 & 14 & 14 & 10 & 8 \\ 16 & 0 & 27 & 18 & 23 & 28 \\ 13 & 30 & 0 & 19 & 27 & 43 \\ 24 & 16 & 15 & 0 & 12 & 17 \\ 12 & 20 & 43 & 23 & 0 & 12 \\ 19 & 42 & 30 & 39 & 18 & 0 \end{array}\right),

(In this case, n_i=4 so we ignore the 1/n_i factor.)

In his paper, Keener proposed a ranking method where the ranking vector {\bf r} is proportional to its score. The score is expressed as a matrix product A{\bf r}, where A is a square preference matrix. In other words, there is a constant \rho>0 such that s_i=\rho r_i, for each i. This is the same as saying A {\bf r} = \rho {\bf r}.

The Frobenius-Perron theorem implies that S has an eigenvector {\bf r}=(r_1,r_2,r_3,r_4,r_5,r_6) having positive entries associated to the largest eigenvalue $\lambda_{max}$ of A, which has (geometric) multiplicity 1. Indeed, A has maximum eigenvalue \lambda_{max}= 110.0385..., of multiplicity 1, with eigenvector

{\bf r}=(1, 1.8313\dots , 2.1548\dots , 1.3177\dots , 1.8015\dots , 2.2208\dots ).

Therefore the teams, according to Kenner’s method, are ranked,

Army < Lafayette < Lehigh < Bucknell < Holy Cross < Navy.

This gives a prediction failure rate of just 6.7\%.

Memories of TS Michael, by Thomas Quint

TS Michael passed away on November 22, 2016, from cancer. I will miss him as a colleague and a kind, wise soul. Tom Quint has kindly allowed me to post these reminiscences that he wrote up.


Well, I guess I could start with the reason TS and I met in the first place. I was a postdoc at USNA in about 1991 and pretty impressed with myself. So when USNA offered to continue my postdoc for two more years (rather than give me a tenure track position), I turned it down. Smartest move I ever made, because TS got the position and so we got to know each other.

 

We started working w each other one day when we both attended a talk on “sphere of influence graphs”. I found the subject moderately interesting, but he came into my office all excited, and I couldn’t get rid of him — wouldn’t leave until we had developed a few research ideas.
 

Interestingly, his position at USNA turned into a tenure track, while mine didn’t. It wasn’t until 1996 that I got my position at U of Nevada.
 

Work sessions with him always followed the same pattern. As you may or may not know, T.S. a) refused to fly in airplanes, and b) didn’t drive. Living across the country from each other, the only way we could work together face-to-face was: once each summer I would fly out to the east coast to visit my parents, borrow one of their cars for a week, and drive down to Annapolis. First thing we’d do is go to Whole Foods, where he would load up my car with food and other supplies, enough to last at least a few months. My reward was that he always bought me the biggest package of nigiri sushi we could find — not cheap at Whole Foods!
 

It was fun, even though I had to suffer through eight million stories about the USNA Water Polo Team.
 

Oh yes, and he used to encourage me to sneak into one of the USNA gyms to work out. We figured that no one would notice if I wore my Nevada sweats (our color is also dark blue, and the pants also had a big “N” on them). It worked.
 

Truth be told, TS didn’t really have a family of his own, so I think he considered the mids as his family. He cared deeply about them (with bonus points if you were a math major or a water polo player :-).
 

One more TS anecdote, complete with photo.  Specifically, TS was especially thrilled to find out that we had named our firstborn son Theodore Saul Quint.  Naturally, TS took to calling him “Little TS”.  At any rate, the photo below is of “Big TS” holding “Little TS”, some time in the Fall of 2000.

tslittlets_photo2000

TS Michael in 2000.