You are on page 1of 21

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/324721039

Sports Analytics: Kabaddi

Presentation · December 2017


DOI: 10.13140/RG.2.2.23850.52164

CITATIONS READS

0 1,521

1 author:

Manojkumar Parmar
Bosch
28 PUBLICATIONS   43 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Kabaddi Analytics View project

All content following this page was uploaded by Manojkumar Parmar on 24 April 2018.

The user has requested enhancement of the downloaded file.


Sports Analytics: Kabaddi
Manojkumar Parmar
manojkumar.parmar17@iimb.ac.in ;
parmarmanojkumar@gmail.com

Title : KABADDI: FROM AN INTUITIVE TO A QUANTITATIVE


APPROACH FOR ANALYSIS, PREDICTIONS AND STRATEGY

No. : BAI1823 || Track: 10 || Venue: K Cluster, 1st Flr – Room K-21

Date : Tuesday, 12/12/2017


Motivation

• Curiosity
• Urge to contribute back
• Culture
• Hobby project
• Trying out my skills
Types of Sports Analytics
Type Description Status w.r.t. Kabaddi
Players’ Injury • To predict players’ • Fragmented and Piecemeal research available over long
& Health fitness and readiness periods
Analytics for game • Consolidation is necessary to generate real value

Business • To leverage business • Not in Focus of Researchers


Analytics aspects of a game • Recent evidence suggest limited usage
Game & • To help predict • Player performance based on fitness & body structure is
Player individual player available with limitation
Performance performance • Player performance in general is area of further
Analytics • To help predict game research
• Game Performance analytics research is non existent
outcomes

Focus of Study is “Game Outcome Prediction”


Research Method

Predictive
Understanding Model Building
Curation Of Of Dataset • Building multiple
Dataset models
• Descriptive Analysis
• Codebook • Visualization • Choice of best
• Tidyup models
• Hypothesis Testing
• Ensemble
modelling
Dataset
• 3rd Standard Style Kabaddi
World cup tournament - 2016
One of its kind dataset for
• 33 Matches (66 Entries)

• 31 Variables
Kabaddi, published on Kaggle

○ Multiple type Attack Points Platform


○ Multiple type Defense Points ○ Received “Featured” Status

○ Other Points ○ Preprocessing Kernel

○ Result of game & toss ○ Codebook

○ Game stage in league


Complete & Consistent Small DataSet
Descriptive Statistics
& Hypotheses
Descriptive Statistics
Average Points /Match /Team
2
4 18

23

11
5

Tackle Pnts All out Pnts Extra Pnts


Touch Pnts Bonus Pnts Provides Overview of Avg. Points
Team Wise
Visualisation
Insights
• Best Defense : Kenya
• Best Attack : India
1
• Average Ranking 2 3
• India
• Bangladesh
• South Korea
More Visualization
Insights
● Toss Results has no meaning

Insights
● Majority of matches are
won with greater margins
Tournament Level Hypotheses
Team wise Hypotheses Results

Insights
• India follows 4
• Australia follows 0
• USA follows 1

Provides Overall Team Profiling


Model Building
Prerequisite for Model Building

Split of Data
• Model Building is an iterative process
• Model Robustness
• Model Accuracies
• Data Split
• Training & Building Model (60%)
• To improve and tune model (20%)
Training CrossValidation Test • To check accuracy & robustness (20%)

Important Practice for Model Building


Feature Engineering
• Understanding variables and their • Recursive Feature Elimination algorithm
features (uses backward feature selection process)
• Essential first task in Model Building • Configuration of RFE for study
• Helps in identifying appropriate • Random forest as helper function
predictor variables • Cross validation – 5 times
• Demanding process needing • Random samples – 20
profound understanding of data and
domain
• From RFE output, top 5 predictors
selected (out of 21) touchPntsDiff,
• Good news : Feature Engineering can raidPntsDiff, alloutPntsDiff, tacklePntsDiff, and
be automated in limited way alloutPntsRec

RFE top 5 predictors used for building all models


Pure Models Models

Support Vector Support Vector


Logistic Tree Based Neural Network
Machine based Machine based
Regression Based Models based
Linear Kernel Radial Kernel

glm rpart svmLinear 1 svmRadial nnet

glmnet ctree svmLinear 2


• Models designed to provide
binary out come of Win/Loss
• Some Models also provide
rf svmLinear 3
probability of Win/Loss
Model Performances

• Overfitting
• rf
• rpart

• Low
Accuracy
• ctree
Ensemble Modeling
• Technique of combining multiple
models glmnet (100%)

Ensemble Models
• To improve accuracy and robustness
Model 1 (rf) rf (42.21%)
• Selection criterion for base models
• For study, variation in the importance nnnet (0%)
of variables among set of base learners
nnet (100%)
• Refer table 4 in paper
Model 2
• Failed to demonstrate any (glm)
glmnet (3.36%)
performance boost
rf (0%)
Further Research
• Reinforced Learning
• Agent modeling based on predictive models

• Generic validation of
• Environment modeling based on standard
kabaddi rules
hypotheses
• Objective is to maximize score
• Generalisation of Models
• Generate knowledge with artificial
• Better Visualisation Models intelligence with RL
• Field Testing • Unique game strategies
• Offset Role of Coach
• Devising Strategies, Selecting line up, Tactics &
substitution
Key takeaways from Study

Technique Utility Usecase


Violin Plot Visualisation of team behavior Devising winning strategies
Hypotheses Visualisation Visualisation of dominant behavior Deciding strategies & line up for a game
Models Game prediction outcome Deciding tactics & substitution
Model - glmnet Coefficients to link out come Devise strategies & scenarios
Model - nnnet Accurate outcome but difficult to To test impact of various strageies
understand

Caution : Results
generalization can
suffer extreme biases
due to small data size
“ The important thing is not to stop
questioning. Curiosity has its own
reason for existing
- Albert Einstein

Thank You!!
View publication stats

You might also like