Mining geospatial temporal patterns from soccer games

1
Mining geo-spatial temporal patterns from soccer
games

• Paul Lesov G2
• 8715 Spatial Databases
• University of Minnesota
• Fall 2007

2
Outline

• Motivation
• Problem Statement
• Challenges
• Scope/Limitations
• Contribution
• Spatial Indexing
• Relational Model
• Temporal event handling
• Demonstration
• Future Work
• Conclusion

3
Motivation

• Importance
• Soccer is the most popular game in the world
• Soccer has great impact on many regional
  economics
• Soccer data is used for coaching, marketing,
  medical, AI simulations and betting odds
• Finding hidden trends in the data collected can
  lead to improved coaching, injury prevention,
  targeted marketing, better prediction and more
  realistic game simulations
• Goal Provide a database framework to allow
  spatial- autocorrelation of a soccer game, which
  would in turn allow applicable statistical
  methods to find useful patterns and nuggets
  contained within

4
Problem Statement

• Current Data Mining Limitations
• Today trends are data mined from the non-spatial
  aspects of the game (player and team statistics)
• Spatial-centric game play information is obtained
  as needed by reviewing game video
• Data Mining and aggregations of spatial aspects
  over multiple games is not possible
• Example A team owner may be considering
  expanding the size of the pitch. He may want to
  know if the team performance will suffer as a
  result of such move. A question such as During
  last 3 seasons was out team more successful at
  (passing/dribbling/shooting/ scoring) when
  attacking through the middle or the flanks? can
  not be easily answered quantitatively today.

5
Challenges

• Manual data collection is time consuming
• Advances in obtaining spatial data from video
  footage
• GPS device embedded in a ball and in a future
  possibly players clothes
• Spatial-Temporal Aspects
• Lack of defined standard for capturing temporal
  relations between spatial entities
• Implementation Issues
• OGIS extensions to popular free databases are not
  fully standardized and support for different
  functions is varying to a large degree
• Development was done using both MySQL and
  PostgreSQL to allow for an overlaps between the
  two providers to cover all needed functionally

6
Contribution Spatial Indexing

• Key Concepts
• Distance Preserving Fixed Spatial Grid
• well defined, static boundaries
• data is evenly distributed
• column-based ordering for 60 10x10m quadrants
  emphases the importance of horizontal movement
  down/up the field

7
Contribution ER
Spatial Non Temporal
Spatial Temporal
Non Spatial Temporal
8
Contribution Temporal

• Event entity maintains all the temporal aspects
  for a single game. It has no spatial
  representation and is an identifier for spatial
  Entities Pass, Dribble and Shot. A spatial
  Possession entity is a collection of the Events
  spatial derivatives.

INPUT EVENT E, PASS P, SHOT S, DRIBBLE
D OUTPUT POSSESION PO For each E Initialize
id0, Geo_collection_array()j0 If
id.time_end(id).time_start If
p.eventid Add p.geom to geo_collection_array
else if d.eventid Add d.geom to
geo_collection_array else if s.eventid Add
s.geom to geo_collection_array else j copy
geo_collection_array to PO PO.idj update
possesion with PO.id
9
Scope/Limitations

• Manual Data Collection
• Limited data set prevented us from multi-game
  pattern mining
• Only ball trajectory is collected, not 22 players
  and 3 officials
• Limitation on knowing kinematics of each player
  and players spatial relations to each other
• Game Play elements identification of which is not
  supported by database schema
• Fouls
• Injuries
• Set pieces
• Temporal data resides independently of spatial
  possession data
• Similarity identification queries are not
  supported (Such as show all possessions that
  followed the same path as this possession)

10
Validation/Demonstration
A prototype is accessible at http//65.41.192.13/g
soccer.htm It utilizes colors, shapes and
different line types for the resulting visuals.
Example Show all long (over 20 meters) passes
into penalty area between 60th and 70th minute.

11
Validation/Demonstration

Example Show all possessions for team Germany
which resulted in lost of possession from
dribbles into opponents penalty area
12
Future Work

• Automate data collection
• Utilize emerging video stream player
  disambiguation technology or GPS tracking
• Extend database schema to support
• Fouls, Injuries and Set pieces (easy)
• 22 players 3 officials kinematics (harder as
  player actions are not easily identifiable and
  players have complex trajectories temporal
  aspects of which must be preserved )
• Allow similarity identification queries
• Polynomial approximations for most of the
  possessions is possible as curves and usage of
  PA-Tree for indexing

13
Conclusion

• We have created a prototype for storing
  spatial-temporal data contained within a soccer
  game by providing
• ER model
• distance preserving fixed grid
• functional approach to temporal issues
• working demonstration
• Mining this data over the course of many games
  may realize much insight into the game and can
  lead to improved coaching, injury prevention,
  targeted marketing, better prediction and more
  realistic game simulations

14
Questions?
---THANK YOU!