Mining Baseball Statistics

1
Mining Baseball Statistics
Data Mining CSE881
Paul Cornwell Kajal Miyan Mojtaba Solgi Project
URL http//kmp-cse881.appspot.com/

2
Overview of Baseball

• Baseball is a team sport
• There are two major leagues AL (American), NL
  (National)
• Many statistics characterizing player performance
  are published yearly
• Each league names one player MVP (Most Valuable
  Player) each year according to a vote
• People place bets on who will be MVP

2
3
Overview

• Application (motivation)
• Can we predict who will be named MVP?
• Learn how to do data mining
• Learn about baseball
• Impress sabermetricians
• Baseball its not diseases, crime, or pollution
• Baseball statistics
• Main task predict MVPs for a given year
• Use SVM to rank players

3
4
Overview of Data and Mining

• Data 5 CSV files (Batting, Fielding, Master,
  Awards, Salaries)?
• Data Mining
• Ranking (similar to classification)?
• Anomaly detection (maybe)?

4
5
Methodology - Preprocessing

• Initial Data 90,000 rows in Batting table,
  1871-2007
• One row one player/year/stint/team
• Cut to 1985-2007, 28,000 rows, b/c Salary begin,
  rule changes
• Perl script to merge tables by playerID/yearID/sti
  nt
• Batting?Fielding?Awards(MVP)?Salaries?Master
  48 columns
• 14 hours, but I got to relearn Perl!
• Discovered infeasible to use WEKA, need to use
  SVM-Light
• Reformatted from CSV to space-delimited SVM-Light
  format
• replace every value with attributevalue
• replace commas, spaces
• deleted 131 w/out fielding record (3-max 26, 21,
  16 at-bats)?
• create (binary) rank value based on MVP status
• replace all MM/DD/YYYY with YYYY
• insert qid column according to year/league (46
  qids)?
• ...

5
6
Methodology Data Mining

• Classification not apt to get good results, hence
  ranking with?
• SVM-Light (Cornell University)?
• Training generates a model which can rank input
• Training phase Leave one (year) out
• Testing Rank the players for that year
• Postprocessing
• SVM-Light returns only ranks of the players as
  integers
• match ranks with corresponding players
• Reformat data for visualization
• Ranked the data for each attribute
• Anomaly detection (in progress)
• KNN on 4 attributes (Gbat, R, HR, RBI)? for
  players in gt 10 games
• Compute z-scores for each attribute/year
• Rank players by distance from nearest neighbor
• Compare ranks in various attributes for detecting
  anomalies

6
7
Methodology - Visualization

• Bar charts of top 20 ranked players for various
  attributes
• Python
• Google App Engine
• Google Charts tool
• U.S. map of player birthState density

7
8
Team Roles

• Roles of team members
• Planning - Everyone
• Preprocessing Paul Cornwell
• Data Mining Kajal Miyan
• Visualization Mojtaba Solgi

8
9
Related Work

• No apparent academic work on predicting MLB MVPs
• PECOTA
• Baseball Prospectus
• www.baseballprospectus.com/pecota/
• Baseball forecasting
• Makes statistical predictions about players
• No MVP prediction evident
• subscription service
• Books are available with baseball forecasts
• apparently for one year only

9
10
Experimental Setup

• Raw data downloaded from http//baseball1.com/cont
  ent/view/58/82/
• Preprocessing done using Perl, Nano, Excel, OOo,
  TextPad
• Preprocessing yields a table with 28K rows and
  45 columns
• Experiments were conducted on a 2 GHz P4 machine
  running Kubuntu 8.04 with 1GB RAM
• Data Mining and postprocessing with SVM-Light,
  Visual C, Matlab
• Visualization done using Python, Google App

10
11
Experimental Evaluation

• Preliminary results
• SVM-Light trained on 1985-2006 data
• tested on 2007
• ranked actual MVPs 1 and 11 (out of 1242
  players) (2nd NL, 2)?
• (there is one MVP for each league each year AL,
  NL)?
• 2006 ranks 7, 16 (1371 players)
• 2005 ranks 1, 4 (1322 players)
• 2004 ranks 1, 3 (1342 players)
• 2003 ranks 3, 32 (1341 players)
• 2002 ranks 1, 11 (1316 players)
• Final evaluation (pending)?
• Leave-one-out

11
12
Visualization Demo

• http//kmp-cse881.appspot.com/

12
13
Conclusions

• MVP ranking was surprisingly successful
• Early results suggest that it is feasible to
  predict MVPs with some accuracy
• Lessons learned
• Data mining is hard work
• Baseball statistics are actually sort of
  interesting
• Future work
• Leave-one-out validation
• Incorporate team statistics in player evaluations
  (expert advice)?

13