Predicting CrossCountry Results using Feature Selection and Evolutionary Computation

1
Predicting Cross-Country Results using Feature
Selection and Evolutionary Computation

• Caio Soares
• Juan E. Gilbert
• Auburn University Auburn, AL
• 2009 Richard Tapia Celebration of Diversity in
  Computing Conference
• 04/02/2009

2
Agenda

• Cross-Country
• Current Ranking System
• Data
• Approach
• Results
• Conclusions
• Future Work

3
Cross-Country Running
NO
YES
4
Cross-Country Running

• Cross-country consists of a foot race with
  distances ranging from 4K to 12K (roughly 2.5 to
  7.5 miles)
• Varied running terrain
• Various Levels Middle School to Professional
• NAIA (National Association of Intercollegiate
  Athletics)

5
Cross-Country Ranking

• Similar to NCAA Basketball and Football
• Route to National Meet
• Region Winners
• At-large bids (based on Rankings)
• Rankings determined by Coaches
• Bias

6
Data

• Lack of Training Data
• Most race data unavailable and hard to collect
• 2005 2007 Regional Meet results only
• Attributes (Metrics)
• Final Coaches Ranking
• Teams Top Runner Time at Region Meet
• Teams Average Time at Region Meet
• Teams 1-5 Split at Region Meet
• Teams Region
• Teams Region Place
• Teams National Place

7
Accuracy Comparison

• A team receives 1 point for each team it beat and
  it was ranked ahead of, meaning it was supposed
  to beat.
• It receives 1 point for each team it lost to and
  was ranked behind of, meaning it was supposed to
  lose to.
• It receives 0 points for each team in which the
  opposite of either case above happens.

8
Approach

• Feature Extraction
• Feature Selection
• Evolutionary Computation
• Particle Swarm Optimizer (PSO)

9
Feature Extraction

• Transformation or combination to create new
  feature
• Every feature turned into 1 n ranking
• Top Runner 1-5 split
• Region Goodness Ranking (based on number of
  teams sent to Nationals)

10
Feature Selection

• High-dimensional Data
• Curse of Dimensionality
• Done in Exhaustive Manner
• Every combination tried
• Candidate sets compared by point accuracy
• Most accurate subset kept

11
Evolutionary Computation

• Assign weights to features
• Find optimal weight set
• Based on point accuracy
• Based on natural selection (simulated evolution).
• A population of candidate solutions (individuals)
  is randomly generated and each is assigned a
  fitness.
• Parents are selected and offspring created
• Survivors are selected
• Process continues until stopping point is reached

12
Particle Swarm Optimizer (PSO)

• Not based on natural selection, based on social
  interaction of swarm of birds
• Designed to embody concept that social sharing of
  information provides evolutionary advantage

13
Particle Swarm Optimizer (PSO)
Ik lt x-vector gt lt p-vector gt lt v-vector
gt x-fitness p-fitness
Ring-Topology
I1
I2
I5
I3
I4
Star-Topology
vi vi ?1rnd()(pi-xi)
?2rnd()(pg-xi) xi xi vi
I1
I2
I5
I3
I4
14
Particle Swarm Optimizer (PSO)

• swarm size 15, 20, 25, 30, 40
• Vmax ¼ Xmax, ½ Xmax, equal to Xmax
• f1 2.4, 2.6, 2.8, 3.0, 3.2, 3.4
• f2 1.2, 1.4, 1.6, 1.8, 2.0, 2.2
• Clercs Coefficient
• Asynchronous updates

15
Training/Testing

• Lack of data
• 2 Years Training/1 Year Testing
• Done 3 times (2005, 06, 07)

16
Accuracy of isolated features in Mens Rankings
17
Accuracy of isolated features in Womens Rankings
18
Yearly Accuracy in Mens Rankings
19
Results (k values)
20
Yearly Accuracy in Womens Rankings
21
Overall Accuracy (2005 to 2007)
22
Conclusion

• On average, PSO outperforms existent system
  (Coaches Poll)
• System provides fairer and less subjective way of
  selecting teams which will enter the National
  Meet through an at-large bid (No Bias)
• Lack of Data makes it hard to predict accurately

23
Future Work

• Standardize data collection and collect
  throughout season
• Factor in individual athletes, race conditions,
  athlete injuries
• Hybrid model between algorithm and Human
  subjective voting

24
The End
Questions? Comments?