Sage Journals: Discover world-class research

Abstract

This study investigated the potential use of k-means clustering for organising training groups in Australian Football (AF) based upon match performance data. Global positioning system (e.g., distance per minute) and technical action (e.g., kicks) data were collected on 38 AF players from 22 matches during the 2021 season. Each player was assigned their season average for all metrics, as well as their primary playing position (the position they occupied the most during the season). Following the application of the elbow method, the data was grouped into four clusters; cluster one could be defined as the low technical action players, cluster two as the low physical output players, cluster three as the high technical action players, and cluster four as the high physical output players. The results of the k-means clustering revealed that each group had unique match performance characteristics compared to each other. The clustering revealed that the majority of players within each cluster shared common playing positions, similar to traditional AF groupings. However, it was also able to highlight that some players were grouped away from their traditional positions based upon their match performance data, indicating that they may require an alternative training stimulus to successfully prepare for competitive matches. Training recommendations for each of the four groups are provided. The study indicated that k-means clustering could provide a useful additional tool for the organisation of training groups within AF populations, and potentially for other similar team sports.

Keywords

Athlete preparation global positioning system high-velocity running kicking performance analysis training

Introduction

Australian Football (AF) is an intermittent effort sport played between two teams of 18 players, plus four players upon the interchange bench. During competitive matches, players are generally required to cover between 11,000 and 13,500 m whilst also repeatedly performing several technical actions such as kicks, handballs, tackles and marks.^1,2 Previous research highlights the importance of both these physical and technical outputs in relation to successful match performance^3–7. Specifically, Wing et al.⁴ established the importance of tackling and marking during successful phases of defence, with superior running performance noted during successful phases of offence. Additionally, team kicking and goal conversion have been previously highlighted as the most important performance metrics when explaining team success.⁵

The importance these physical and technical outputs have upon team success highlights the need for training that effectively develops AF players running ability and skill proficiency. A comprehensive study by Tribolet et al.⁸ describes the multifaceted nature of AF team training modalities, which range from fundamentals, skill development, line training, match simulation and fitness or conditioning. Additionally, Loader et al.⁹ were able to classify AF training drills into three clusters: game specific conditioning, skill refining moderate intensity dominant, and skill refining low intensity dominant. During skill development drills (e.g., skill development, line training and fundamentals), players are often placed into groups based upon their playing position (e.g., backs, midfielders, and forwards), where training is focused upon position-specific tasks. Although there are undeniable benefits to dividing players into positional groups for the purposes of training, the tactical evolution of AF¹⁰ has seen the growth of hybrid type players, capable of playing within multiple positions.⁷ Additionally, previous clustering of player positions within the Australian Football League Women's (AFLW) competition, revealed multiple player roles that sit within playing positions (e.g., high-disposal defenders) may be present.¹¹ Therefore, it may be beneficial to explore alternative grouping strategies within modern AF training practices. This may be particularly important when training time is at a premium, hence emphasising the most relevant individual traits from competitive matches may optimise training effectiveness within training time constraints.⁵

Utilising a statistical clustering method which can account for multiple metrics, can be seen as beneficial to achieving optimal grouping. Several different clustering methods (e.g., fuzzy, k-median, k-medoid, k-means, hierarchical) can be employed to achieve this, each with their own strengths and weaknesses.^12,13 Specific to sports performance research, clustering methods have been employed in soccer (fuzzy clustering),^14,15 basketball (fuzzy clustering, complex networks, Bayesian networks, two-step clustering),^16–19 American football (k-means),²⁰ tennis (fuzzy, stochastic blockmodel),²¹ and AF (Wards two-way hierarchical, K-means, CLARA).^9,11,22 Which method to use relies upon striking a balance between statistical robustness, ease/burden of computation, and practical application. This particular research uses the k-means clustering method previously utilised by Shelly et al.,²⁰ who have made progress towards establishing the use of k-means for devising training groups in American football. However, further studies are warranted within this area to validate the technique in this context. Additionally, the study by Shelly et al.²⁰ was limited to global positioning system data only, where the inclusion of technical actions (e.g., kicks) may prove beneficial when devising training groups. This is due to the reported benefit of performing these actions at match speed in order to increase positive transfer to performance.^8,23

K-means clustering is an unsupervised machine learning method whereby n observations are partitioned into k clusters (also referred to as centres).²⁴ This results in a set of clusters which are dissimilar to each other, but the observations within each cluster are similar to each other.^20,24 In this instance, an AF player would be an observation, and sit within a cluster where their match performance data are similar to the other players within the same cluster, but different to those players partitioned into the other clusters.

To perform this successfully as it relates to training design, both physical performance (e.g., distance per minute) and technical actions (e.g., kicks) should be included within the data collected. This is not only due to the importance of these metrics to successful performance, but it has also been noted that technical actions should be performed at match speed in order to increase positive transfer from training to match performance.^8,23 This concept is further supported by Farrow et al.²⁵ who demonstrated that generally open drills displayed significantly more game-like decisions and higher cognitive complexity scores, as well as higher physical outputs, as opposed to closed drills. Therefore, the aim of this study was to assess the potential and utility of k-means clustering as a method for aiding the division of training groups for AF players using match performance data.

The remainder of this paper is structured as follows: Section 2 outlines the methodology utilised for the analysis; Section 3 presents the findings; Section 4 discusses the findings in respect to Australian Football training design; Section 5 outlines the limitations and future research recommendations; and Section 6 provides conclusive remarks.

Methods

Participants

Microsensor technology and technical action data were collected on 38 male sub-elite AF players (age: 23.0 ± 3.3 years; body mass: 82.8 ± 10.1 kg; height: 185.8 ± 8.2 cm) from one club competing in the West Australian Football League (WAFL) during the 2021 season across 22 matches. Individual player observations were removed if the player was unable to complete the entire match (e.g., sustained an injury), or if there was failure of the recording device.²³ This missing data is considered to be missing completely at random (MCAR). A total of 454 player observations (average observations per player 11.9 ± 6.9; range 2–22) were included in the final analyses. Additionally, players were split into 7 positional groups according to the position they spent the most amount of on-ground playing time in each individual match. This data was obtained from the interchange software which provides the time played in each position (Interchanger, Australia). They were then assigned the position they played the most matches in as their final position for the purposes of this study. This included full back (n = 4), half back (n = 8), inside midfield (n = 10), wing (n = 4), half forward (n = 5), full forward (n = 5), and ruck (n = 2). This study was approved by the Edith Cowan University Human Research Ethics Committee (HREC ID: WING-2023–04977).

Data collection

Microsensor technology data was collected using the Playertek device (Catapult Innovations, Melbourne, Australia) sampling at 10 Hz. The accuracy of these devices has been previously reported.²⁶ To reduce interunit variability, all players wore the same unit for each match, fitted into a specifically designed pouch sewed into the playing shirt. Upon completion of each match, the data was downloaded onto the proprietary software (Playertek Cloud), where the start and end times of each quarter were synced from the Playertek live-feed application. All time spent upon the interchange bench was removed so that the analyses was formed of on-ground time only. The following metrics were collected from the microsensor device and expressed per minute of playing time for the entire match (i.e., the sum of the four quarters); total distance (m), high-speed distance (HSR; >18 km h⁻¹), very high-speed distance (VHSR; >24 km h⁻¹), PlayerLoad^TM (the sum of accelerations across the three accelerometer vectors accounting for the instantaneous rate of acceleration and divided by a scaling factor,²⁷ PL; AU), accelerations (efforts >3 m s⁻²), and high-intensity efforts (HIE;>18 km h⁻¹ for ≥ 2s duration). These metrics were selected as they have been commonly reported within the AF literature^4,23,28–31

Additionally, technical action data was acquired from Champion Data (Melbourne, Australia), a company that provides statistics to the Australian Football League and the WAFL. Their coding of events has been previously shown to demonstrate acceptable levels of accuracy.³² These technical actions included³³:

Kick: Disposing of the ball by any part of the leg below the knee.

Handball: Disposing of the ball by hand.

Tackle: Preventing a successful disposal by on opponent through using physical contact.

Mark: Catching a kicked ball that has travelled >15m that has not been touched by another player or the ground.

These technical actions were included as they are both commonly highlighted within the AF literature, and have been reported to be important to successful match performance.^1,4,23,32 Each player was assigned a season match average for all microsensor and technical action data within the final data set. Utilising a combination of data modalities (e.g., microsensor technology and match statistics), can provide a greater spectrum of data to understand player performance.³⁴

Statistical analyses

All statistical analyses were performed in Microsoft Excel (Microsoft Corporation, Redmond, WA, USA), and R (R for Statistical Computing, version 4.3.2)³⁵ where the packages “factoextra”³⁶ and “cluster”³⁷ were utilised. To ensure all data points contributed equally to the model, the data was first scaled. This was performed by using the scale function within R, which first centres the data by subtracting the mean, and then scales it through dividing by the standard deviation. To determine the optimal number of clusters, the elbow method was utilised through producing a scree plot where the total within-cluster sum of squares was plotted against each candidate of k clusters.²⁴ The point at which the graph represents an “elbow” corresponds to the optimal number of clusters (Figure 1), where there is a trade-off between both minimizing the number of clusters and the variance within each cluster.²⁴ The scaled data was then passed into the k-means algorithm, where the optimal number of clusters (four) was set from the result of the scree plot analysis. Group descriptives (average and SD) were then recorded for each of the metrics for each cluster, as well as the playing positions of the players in each cluster.

Figure 1.

Scree plot of total within sum of squares plotted against number of clusters (1 to 10), utilised as the “elbow method” for ascertaining optimal number of clusters used within the final analyses.

Following the application of K-means clustering, silhouette coefficients were extracted to assess clustering performance. Additionally, divisive hierarchical clustering (DIANA) analysis was also performed post clustering to further assess clustering performance.

Results

The results of the elbow method demonstrated the optimal number of clusters to be four, which are depicted in Figure 2, where each coloured shape represents a cluster, and each number a player within that cluster. The post clustering silhouette coefficients yielded an average score of 0.22, with a positive score deemed acceptable. The k-means clustering was performed again at all iterations of k from two to thirteen (selected as thirteen player roles were identified by van der Vegt et al.¹¹), with the silhouette coefficient average score ranging from 0.20 to 0.26, indicating that the silhouette score was not drastically improved at any number of k. Furthermore, DIANA analysis was also performed, with the results suggesting the presence of four clusters. The results of the post clustering analysis, combined with the elbow method and our domain knowledge of AF training (e.g., the number of player groups that may be practically possible to employ during team training),³⁸ satisfied our interpretation of four clusters being present within the data set.

Figure 2.

Cluster plot revealing the results of the K-means clustering using four clusters (centres) as derived from the scree plot (figure 1). Each cluster represents a group of players, characterised by specific match performance characteristics. Cluster one; low technical action players, cluster two; low physical output players, cluster three; high technical action players, cluster four; high physical output players. Each number within each cluster represents an individual player. Dim 1 and Dim 2 represent the first two principal components, with Dim 1 accounting for the greatest variability in the data set (48.7%), and Dim 2 the next highest (20.9%) whilst being orthogonal to Dim 1²⁰.

Finally, the k-means clustering was checked for sensitivity by performing a “leave-one-out” cross validation, where the k-means was re-performed 38 more times with one player systematically left out on each occasion, with the clustering results compared to those obtained utilising the full data set. This cross validation demonstrated that 31 players remained in the same cluster on all 38 occasions. Five players remained in the same cluster on 37/38 occasions, one player remained in the same cluster on 31/38 occasions, with one player remaining in the same cluster on 22/38 occasions.

Each of the clusters can be broadly classified, with cluster one containing players who recorded lower values for handballs (vs cluster three), kicks (vs cluster three), and tackles (vs cluster three and four). Cluster two represents the low physical output players (e.g., those who recorded the lowest microsensor technology values, predominately for distance and high-speed running). Cluster three represents the high technical action players (e.g., those who recorded the highest values for kicks and handballs). Cluster four represents the high physical output players (e.g., those who recorded the highest values for the microsensor technology metrics, predominately for distance and high-speed running). The descriptives for each cluster are shown within Table 1.

Table 1.

Descriptives (average ± SD) for microsensor technology metrics, technical actions, and playing positions for the four clusters.

Cluster	Microsensor Technology Metrics						Technical Actions				Playing Position
Cluster	DIST	HSR	HIE	PL	VHSR	ACCEL	Kicks	Handballs	Marks	Tackles	FB	HB	M	W	R	HF	FF	Total
1	128.4 ± 6.6	23.1 ± 3.1	0.7 ± 0.1	5.5 ± 0.4	4.4 ± 0.8	0.7 ± 0.1	7.2 ± 2.2	4.3 ± 1.3	3.3 ± 0.9	2.0 ± 0.8	2	5	0	1	0	4	2	14
2	118.8 ± 4.3	14.3 ± 3.3	0.4 ± 0.1	4.9 ± 0.3	1.7 ± 0.8	0.6 ± 0.1	8.0 ± 2.3	4.3 ± 1.1	3.7 ± 1.9	2.3 ± 1.0	2	1	1	0	2	0	3	9
3	130.4 ± 8.4	21.1 ± 4.8	0.6 ± 0.1	5.6 ± 0.6	3.0 ± 1.0	0.8 ± 0.1	12.2 ± 1.7	8.4 ± 2.1	3.7 ± 0.9	4.0 ± 1.3	0	2	7	0	0	0	0	9
4	149.6 ± 5.5	32.3 ± 3.6	0.9 ± 0.1	6.5 ± 0.7	4.5 ± 1.0	1.0 ± 0.1	7.5 ± 1.8	6.0 ± 1.3	2.9 ± 1.2	4.1 ± 2.4	0	0	2	3	0	1	0	6

DIST: distance/min, HSR: high-speed running/min (>18 km/h), HIE: High-intensity efforts/min, PL: PlayerLoad/min, VHSR: very high-speed running/min (>24 km/h), ACCEL: accelerations/min (>3 m/s/s), FB: Full back, HB: Half back, M: Inside-midfield, W: Wing, R: Ruck, HF: Half forward, FF: Full forward.

Cluster one contained 14 players, of which two were full backs, five half backs, one wing, four half forwards, and two full forwards. The players were generally defined as the lowest for the technical actions, but high for high-speed (23.1 ± 3.1 m min⁻¹) and very high-speed distances (4.4 ± 0.8 m min⁻¹), as well as high-intensity efforts (0.7 ± 0.1 efforts per minute). Cluster two contained nine players, of which two were full backs, one half back, one inside midfield, two rucks, and three full forwards. These players were the lowest for all microsensor technology metrics, but highest for marks (3.7 ± 1.9). Cluster three contained nine players, of which two were half backs, and seven inside midfielders. These players were high for technical actions, and in-particular kicks (12.2 ± 1.7), and handballs (8.4 ± 2.1). Finally, cluster four contained six players, of which two were inside midfielders, three wing, and one half forward. These players were the highest for all microsensor technology metrics and tackles (4.1 ± 2.4).

Discussion

The utility and usefulness of k-means clustering as a method for devising training groups in a population of AF players were assessed. The results demonstrated that this population of sub-elite AF players could be divided statistically into four groups, each with a unique identity in terms of match performance data, which can subsequently be targeted with specific training interventions. It should be noted that the number of clusters derived from the elbow method may differ by team, and be influenced by playing style, the mix of player profiles, and the specific metrics being assessed.²⁰ Although there was some evidence of grouping in terms of playing positions, the players within each cluster were not completely unique to a particular playing position, which is in line with those reported in American football.²⁰ This is perhaps an unsurprising finding amongst AF populations, given the tactical evolution of game play has led to more hybrid-type players, who often switch positions, sometimes even during the same game as well as across the season.^7,10 Therefore, it may be that technical and physical match performance is more person and role specific, as opposed to position specific in modern AF.

Cluster one was defined by the presence of low technical action players, recording the lowest values for kicks (7.2 ± 2.2), handballs (4.3 ± 1.3), tackles (2.0 ± 0.8), and the second lowest for marks (3.3 ± 0.9). However, they did display some of the highest values for the high velocity microsensor technology metrics; HSR (23.1 ± 3.1 m min⁻¹), VHSR (4.4 ± 0.8 m min⁻¹), and HIE (0.7 ± 0.1 efforts per min). This cluster also contained five of the eight (63%) half-backs, and four of the five (80%) half-forwards within the study, who together made up 64% of the 14 players within the cluster. Although it appears counterintuitive to cluster together half forwards and half backs, this is common practice within research, where these two positions are combined to form a half-line grouping.^1,23,39,40 Oftentimes, this occurs because these players play directly against each other in competitive matches and therefore likely to display similar running demands and movement patterns. Interestingly, in our study, these positions were still predominantly clustered together when technical actions are also included, potentially due to the low number recorded. Additionally, three of the five players who were not classified as half-line players within this cluster, had played some games during the season in a half-line position. This strengthens the theory of grouping players by match performance data, as opposed to playing position.

Cluster two was defined by low physical output players, recording the lowest values for all microsensor technology variables. However, this group did record the highest number of marks (3.7 ± 1.9), and the second highest number of kicks (8.0 ± 2.3) per game. This cluster consisted of two full backs, both ruck position players, and three full forwards (50%, 100%, and 60% of the study population respectively), which when combined accounted for 78% of the cluster's population. The clustering together of these positions is somewhat unsurprising considering that these positions are often grouped together and referred to as fixed, full, tall or key position players in AF research.^23,29,39,40 Previous research delineating AF populations into discrete positional groups have noted these playing positions to record the smallest running distances and distances recorded at high velocities.²⁸ This can be attributed to several factors, which include their role within the team requiring them to play in a more restricted area of the playing oval (e.g., forward or defensive 50), they are typically larger in stature, and have been reported to record the lowest scores on physical performance measures.⁴¹ However, their large stature often provides them with an advantage when marking the ball, which is highlighted in the high number of marks recorded by players in this cluster.

Maybe somewhat surprisingly, the other two players within this cluster are one half back and one inside-midfield position players, who would have potentially been expected to have been clustered into other groups. However, it should be noted that the half back played nine of their 20 games as a full back (with the other 11 at half back), which may have led to the player being clustered to this group. Additionally, it may be that their specific role within the team produces an output (physically and technically) that more closely resembles a key position player. For example, the inside-midfielder in this cluster may have played more of a role as the protector, where his job is to block and create space for other midfielders to receive the ball.⁴² In this instance it is likely that the player would have less kicks and handballs (in comparison to the other midfielders), and potentially cover smaller running distances. However, without deidentifying the data and receiving the technical/tactical expertise of the coaching staff, this is somewhat difficult to confirm.

Cluster three was defined by high technical action players, with this group recording the highest number of kicks (12.2 ± 1.7) and handballs (8.4 ± 2.1), as well as recording high values for marks (3.7 ± 0.9) and tackles (4.0 ± 1.3). These players also recorded high values for all microsensor technology metrics. The cluster was formed of seven inside-midfielders (70% of study population), and two half backs (25% of study population). The grouping together of the majority (7 out of 10) of the inside-midfielders within the study is encouraging and strengthens the validity of the methodology. Typically, these players have the most possessions and disposals during an AF match,⁴³ as well as covering large running distance,²⁸ which is mirrored within the descriptives for cluster three. Similarly, the two half backs within this cluster recorded more average disposals (kicks and handballs combined) per match than any of the other half backs within the study, which may be the predominant reason they have been grouped into cluster three.

Cluster four was defined by high physical output players, recording the highest values for all microsensor technology metrics. This group also recorded the highest tackles per game (4.1 ± 2.4). The cluster was formed of three of the four wing position players, two inside-midfielders and one half forward (75%, 20%, and 20% of the study population respectively). Similar to cluster three, it is encouraging that the majority of the wing position players were clustered together. Previous research has reported wing position players to cover the largest running distances in competitive matches,⁴⁴ which is in agreement with the descriptives of cluster four. The combination of high number of tackles, low number of the other technical actions (i.e., kicks, handballs, and marks), and high physical output suggests that the primary role of the players in this group may be to perform a high number of defensive pressure acts.¹ However, additional data is required to confirm this hypothesis.

The findings of the present study support the potential use of k-means clustering as one valuable method to accurately devise training groups of shared characteristics within AF populations. Similar to American football,²⁰ the clusters contained several players from the same playing position or group; however, there appears to be some subtle differences to traditional training groupings which may add value within the practical setting for some aspects of training. For example, grouping together 70% (7/10) of the inside midfielders aligns with common AF training groups, however, the other three inside midfielders were assigned to different clusters, which may provide novel insights for training prescription. Although separating three players from a positional group to perform a different task for part of training may feel counterintuitive, coaches should feel confident that by using the match as a reference point to derive training prescription, a higher level of specificity can be achieved. Using the inside-midfield group as an example, if physical training is prescribed at the intensity of the 7 grouped into cluster three (distance; 130.4 ± 8.4 m min⁻¹, high-speed; 21.1 ± 4.1 m min⁻¹), then this is unlikely to provide a sufficient stimulus for the 2 inside-midfielders grouped into cluster four. These two players recorded a season average for distance 147.3 m min⁻¹ and 145.7 m min⁻¹, and high-speed 31.6 m min⁻¹ and 30.2 m min⁻¹ respectively. Therefore, ensuring a part of training replicates these physical demands, coaches can have a greater level of confidence that players are both physically prepared for competition and the successful transfer of technical actions can be potentially increased.

It is important to note that grouping players based upon the clustering results may not achieve optimal outcomes for all aspects of a training session/week, but would be better utilised within specific scenarios. As described by Tribolet et al.,⁸ AF training can be separated into two broad categories, training-form (fundamentals, fitness, skills), and playing-form (breakdown, transition, line-specific, match simulation). The groups derived by k-means may be better utilised during training-form drills, where players are often divided into small groups and perform repetitive type practice.⁸ Traditionally, players would often perform these drills with players of a similar playing position, however, this may not truly represent the nature in which they play. Recently, van der Vegt et al.¹¹ established that up to thirteen playing roles may be present within a population of AFLW players. The nuance of these roles should be captured for the purposes of training prescription.¹¹ An example of this could be seen within the defensive-mid role (a defender playing a role that is partly representative of a midfielder).¹¹ This group of players will likely need to develop skill under different constraints, most namely time in possession and heightened opposition pressure,^45,46 compared to those they will likely experience in a traditional backs (defenders) grouping. To achieve this, grouping them with similar players who can create such an environment during training-form type training would appear advantageous. Conversely, playing-form type training may be better served with players divided into traditional groupings, so that team cohesion and tactical practice can be optimised, whilst also exposing players to the vast dynamics of AF match play.

To assist coaches and players to utilise this information in AF practice, potential training prescriptions and areas of focus based upon the cluster results are presented in Table 2. The specific targeting of both these physical and technical attributes can maximise both training time and outcomes that are specific to the way in which the athlete performs during competitive matches. Additionally, ensuring technical actions (e.g., kicks and handballs) are performed in training at game pace has the potential for greater transfer to competitive matches.^8,23 Therefore, it may be prudent for those in cluster four to practice not only tackle technique, but also perform these combined during high intensity running drills, as the players within this cluster recorded the highest values for both tackling and microsensor technology metrics.

Table 2.

Training recommendations based upon the cluster results.

Cluster	Areas of focus	Training recommendation examples
Cluster One (Low technical action players)	High velocity running	Conditioning at high % MAS Repeat sprint drills Maximum velocity exposure
Cluster Two (Low physical output players)	Marking Kicking	Lane kicking Marking, contested and uncontested
Cluster Three (High technical action players)	All technical actions (kicks, handballs, marks, tackles) Running distances and accelerations per minute	Lane kicking, handball grids, tackle grids, chaotic handball games, ball movement drills Small sided games MAS, Fartlek
Cluster Four (High physical output players)	Physical conditioning Tackling	Tackle grids, tackling during high velocity running MAS, repeat sprint drills, Fartlek

MAS: Maximum aerobic speed.

Limitations and future recommendations

This study is limited by a few factors which include the number of technical actions utilised within the data set. Including actions such as pressure acts, inside and rebound 50's, and score involvements may have helped describe a player's role more accurately. Additionally, it may have been valuable to include physical performance measures such as jump height, lower body strength, and aerobic capacity (e.g., time trial), which could help to guide the development of specific physical qualities for each cluster of players. Future research may also wish to obtain tactical information from the coaching staff²⁰ to better understand the specific role the player has been asked to play for the team. Additionally, it may be interesting to compare the results of the k-means clustering to the coach's perception of which players have similar traits. As previously mentioned, there are several clustering methods available to practitioners, therefore, future studies may wish to explore these as alternatives to the k-means method.

Conclusion

The use of k-means clustering can provide some novel insights into training groupings within AF players. Through its use, players can be clustered with other players who have similar match outputs, and training can be tailored to specifically improve upon these. It is important to note, that although there appears to be some value in the use of this methodology, it is not intended to be a standalone method for training organisation. Moreover, it can be used as an additional tool to provide an extra layer of detail concerning some aspects of training. For example, in Table 2, we have recommended a focus on physical conditioning for players within cluster four, but not for those in cluster two. Whilst all players are required to be physically fit, including those in cluster two, it appears to be more prudent for those in cluster four, and therefore an additional training stimulus may be required. Finally, the results of this study, combined with those reported in American football,²⁰ support the notion that k-means clustering could be successfully employed across multiple team sports of a similar nature to that of Australian football.

Footnotes

Acknowledgements

The authors wish to acknowledge the athletes and coaching staff of the football club for their contribution to this research;and to acknowledge Champion Data for their permission to use and report this data in the context of this research.

Data availability

Due to the agreement with the football club,the raw data set associated with this manuscript is not available publicly,so supporting data is not available.

Declarations

CW,NHH,FM,SCF,and KN have no conflicts of interest to declare.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research,authorship,and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research,authorship,and/or publication of this article: CW was supported by an Edith Cowan University Higher Degree by Research Scholarship during the data collection period. NHH is supported by an NHMRC Investigator Fellowship (APP2017080).

ORCID iDs

Christopher Wing

Fadi Ma’ayah

References

Johnston

Black

Harrison

, et al. Applied sport science of Australian football: a systematic review. Sports Med 2018; 48: 1673–1694.

Wing

Hart

McCaskie

, et al. Running performance of male versus female players in Australian football matches: a systematic review. Sports Med Open 2021; 7: 96.

Hiscock

Dawson

Heasman

, et al. Game movements and player performance in the Australian football league. Int J Perform Anal Sport 2012; 12: 531–545.

Wing

Hart

Ma’ayah

, et al. Physical and technical demands of offence, defence, and contested phases of play in Australian football. BMC Sports Sci Med Rehabil 2022; 14: 33.

Robertson

Back

Bartlett

. Explaining match outcome in elite Australian rules football using team performance indicators. J Sports Sci 2016; 34: 637–644.

Sullivan

Bilsborough

Cianciosi

, et al. Match score affects activity profile and skill performance in professional Australian football players. J Sci Med Sport 2014; 17: 326–331.

Ryan

Coutts

Hocking

, et al. Factors affecting match running performance in professional Australian football. Int J Sports Physiol Perform 2017; 12: 1199–1204.

Tribolet

Sheehan

Novak

, et al. How does practice change across the season? A descriptive study of the training structures and practice activities implemented by a professional Australian football team. Int J Sports Sci Coach 2022; 17: 63–72.

Loader

Montgomery

Williams

, et al. Classifying training drills based on movement demands in Australian football. Int J Sports Sci Coach 2012; 7: 57–67.

10.

Woods

Robertson

Collier

. Evolution of game-play in the Australian football league from 2001 to 2015. J Sports Sci 2017; 35: 1879–1887.

11.

van der Vegt

Gepp

Keogh

, et al. An investigation of data-driven player positional roles within the Australian football league women’s competition using technical skill match-play data. Int J Sports Sci Coach 2024; 19: 1130–1142.

12.

D’Urso

De Giovanni

Swartz

. Editorial: big data and data science in sport. Ann Oper Res 2023; 325: 1–7.

13.

Saxena

Mohan Goyal

Mittal

. Comparative analysis of clustering methods. Int J Comput Appl 2015; 118: 30–35.

14.

D’Urso

De Giovanni

Vitale

. A robust method for clustering football players with mixed attributes. Ann Oper Res 2023; 325: 9–36.

15.

Cerqueti

D’Urso

De Giovanni

, et al. INGARCH-based fuzzy clustering of count time series with a football application. MLWA 2022; 10: 100417.

16.

Chessa

D’Urso

De Giovanni

, et al. Complex networks for community detection of basketball players. Ann Oper Res 2023; 325: 363–389.

17.

D’Urso

De Giovanni

Vitale

. A Bayesian network to analyse basketball players’ performances: a multivariate copula-based approach. Ann Oper Res 2023; 325: 419–440.

18.

Mateus

Esteves

Gonçalves

, et al. Clustering performance in the European basketball according to players’ characteristics and contextual variables. Int J Sports Sci Coach 2020; 15: 405–411.

19.

Zhang

Lorenzo

Gómez

, et al. Clustering performances in the NBA according to players’ anthropometric attributes and playing experience. J Sports Sci 2018; 36: 2511–2520.

20.

Shelly

Burch

Tian

, et al. Using K-means clustering to create training groups for elite American football student-athletes based on game demands. IJKSS; 2020; 8: 47–63.

21.

D’Urso

De Giovanni

Federico

, et al. Network and attribute-based clustering of tennis players and tournaments. Comput Stat 2024.

22.

Born

Mundt

Mian

, et al. Clustering offensive strategies in Australian-rules football using social network analysis. Information 2024; 15: 364.

23.

Wing

Hart

Ma’ayah

, et al. Physical and technical demands of Australian football: an analysis of maximum ball in play periods. BMC Sports Sci Med Rehabil 2022; 14: 15.

24.

Steinley

. K-means clustering: a half-century synthesis. Bri J Math Stat Psychol 2006; 59: 1–34.

25.

Farrow

Pyne

Gabbett

. Skill and physiological demands of open and closed training drills in Australian football. Int J Sports Sci Coach 2008; 3: 489–499.

26.

Mooney

Malone

Izri

, et al. The running performance of elite U20 Gaelic football match-play. Sport Sci Health 2021; 17: 771–779.

27.

Boyd

Ball

Aughey

. The reliability of minimaxx accelerometers for measuring physical activity in Australian football. Int J Sports Physiol Perform 2011; 6: 311–321.

28.

Coutts

Kempton

Sullivan

, et al. Metabolic power and energetic costs of professional Australian football match-play. J Sci Med Sport 2015; 18: 219–224.

29.

Johnston

Murray

Austin

, et al. Peak movement and technical demands of professional Australian football competition. J Strength Cond Res 2021; 35: 2818–2823.

30.

Edwards

Piggott

Joyce

, et al. The relationship between two measures of physical capacity and match performance in semi-professional Australian rules football. JASC 2015; 23: 117–120.

31.

Delaney

Thornton

Burgess

, et al. Duration-specific running intensities of Australian football match-play. J Sci Med Sport 2017; 20: 689–694.

32.

Robertson

Gupta

McIntosh

. A method to assess the influence of individual player performance distribution on match outcome in team sports. J Sports Sci 2016; 34: 1893–1900.

33.

Australian Football League. Every stat explained, https://www.afl.com.au/news/144837/stats-glossary-every-stat-explained (2017, accessed April 01 2021).

34.

Hynes

O’Grady

O’Hare

. Towards accessible technologies for coaching. Int J Sports Sci Coach 2013; 8: 105–114.

35.

R Core Team. R: A language and environment for statistical computing. R foundation for statistical computing 2023.

36.

Kassambara

Mundt

. Factoextra: Extract and visualize the results of multivariate data analyses, https://CRAN.R-project.org/package=factoextra (2020, 20th December 2023).

37.

Maechler

Rousseeuw

Strufy

, et al. Cluster: Cluster analysis basics and extensions, https://CRAN.r-project.org/package=cluster (2022, 20th December 2023).

38.

Jasinska-Piadlo

Bond

Biglarbeigi

, et al. Data-driven versus a domain-led approach to k-means clustering on an open heart failure dataset. Int J Data Sci Anal 2023; 15: 49–66.

39.

Black

Gabbett

Johnston

, et al. The influence of physical qualities on activity profiles of female Australian football match play. Int J Sports Physiol Perform 2018; 13: 524–529.

40.

Black

Gabbett

Johnston

, et al. Physical fitness and peak running periods during female Australian football match-play. Sci Med Footb 2018; 2: 246–251.

41.

Pyne

Gardner

Sheehan

, et al. Positional differences in fitness and anthropometric characteristics in Australian football. J Sci Med Sport 2006; 9: 143–150.

42.

Australian Football League. Ruck strategy: The centre bounce, https://www.afl.com.au/news/538341/ruck-strategy-the-centre-bounce (2003, accessed 5th April 2024).

43.

Woods

Veale

Fransen

, et al. Classification of playing position in elite junior Australian football using technical skill indicators. J Sports Sci 2018; 36: 97–103.

44.

Veale

Pearce

Carlson

. Player movement patterns in an elite junior Australian rules football team: an exploratory study. J Sports Sci Med 2007; 6: 254–260.

45.

Ireland

Dawson

Peeling

, et al. Do we train how we play? Investigating skill patterns in Australian football. Sci Med Footb 2019; 3: 265–274.

46.

Browne

Woods

Sweeting

, et al. Applications of a working framework for the measurement of representative learning design in Australian football. PLoS One 2020; 15. doi:https://doi.org/10.1371/journal.pone.0242336

The use of K-means clustering for the organisation of training groups in Australian football

Abstract

Keywords

Introduction

Methods

Participants

Data collection

Statistical analyses

Results

Discussion

Limitations and future recommendations

Conclusion

Footnotes

Acknowledgements

Data availability

Declarations

Declaration of conflicting interests

Funding

ORCID iDs

References