TY - JOUR

T1 - Classification of playing position in elite junior Australian football using technical skill indicators

AU - Woods, Carl T.

AU - Veale, James

AU - Fransen, Job

AU - Robertson, Sam

AU - Collier, Neil

PY - 2018/1/2

Y1 - 2018/1/2

N2 - ​In team sport, classifying playing position based on a players’ expressed skill sets can provide a guide to talent identification by enabling the recognition of performance attributes relative to playing position. Here, elite junior Australian football players were a priori classified into 1 of 4 common playing positions; forward, midfield, defence, and ruck. Three analysis approaches were used to assess the extent to which 12 in-game skill performance indicators could classify playing position. These were a linear discriminant analysis (LDA), random forest, and a PART decision list. The LDA produced classification accuracy of 56.8%, with class errors ranging from 19.6% (midfielders) to 75.0% (ruck). The random forest model performed at a slightly worse level (51.62%), with class errors ranging from 27.8% (midfielders) to 100% (ruck). The decision list revealed 6 rules capable of classifying playing position at accuracy of 70.1%, with class errors ranging from 14.4% (midfielders) to 100% (ruck). Although the PART decision list produced the greatest relative classification accuracy, the technical skill indicators reported were generally unable to accurately classify players according to their position using the 3 analysis approaches. This player homogeneity may complicate recruitment by constraining talent recruiter’s ability to objectively recognise distinctive positional attributes.

AB - ​In team sport, classifying playing position based on a players’ expressed skill sets can provide a guide to talent identification by enabling the recognition of performance attributes relative to playing position. Here, elite junior Australian football players were a priori classified into 1 of 4 common playing positions; forward, midfield, defence, and ruck. Three analysis approaches were used to assess the extent to which 12 in-game skill performance indicators could classify playing position. These were a linear discriminant analysis (LDA), random forest, and a PART decision list. The LDA produced classification accuracy of 56.8%, with class errors ranging from 19.6% (midfielders) to 75.0% (ruck). The random forest model performed at a slightly worse level (51.62%), with class errors ranging from 27.8% (midfielders) to 100% (ruck). The decision list revealed 6 rules capable of classifying playing position at accuracy of 70.1%, with class errors ranging from 14.4% (midfielders) to 100% (ruck). Although the PART decision list produced the greatest relative classification accuracy, the technical skill indicators reported were generally unable to accurately classify players according to their position using the 3 analysis approaches. This player homogeneity may complicate recruitment by constraining talent recruiter’s ability to objectively recognise distinctive positional attributes.

KW - discriminant analysis

KW - machine learning

KW - Performance analysis

KW - random forest

KW - rule induction

KW - Physical education and sports

UR - http://www.scopus.com/inward/record.url?scp=85010693229&partnerID=8YFLogxK

U2 - 10.1080/02640414.2017.1282621

DO - 10.1080/02640414.2017.1282621

M3 - Journal articles

C2 - 28125339

AN - SCOPUS:85010693229

VL - 36

SP - 97

EP - 103

JO - Journal of Sports Sciences

JF - Journal of Sports Sciences

SN - 0264-0414

IS - 1

ER -