An Investigation In To The Variables That Make a Fighter Successful in the Ultimate Fighting Championship
Dataset from kaggle (https://www.kaggle.com/rajeevw/ufcdata) combined with web scraping of sherdog website.
Statistical analysis of sports data to understand important variables for match predictions has been a study of the machine learning literature in the last 10 years[1], [2].
Predominantly focused on well established sports, more niche sports like mixed martial arts (MMA) have been largely ignored [3]. MMA presents an interesting problem, as a fighters skill set appears to be an important predictor of the winner of a fight (28 current champions are wrestlers, whilst 6 are kickboxers). All martial arts specialise in one method of winning either knockouts (from standing position) or submissions (from ground position). As both are permitted in MMA, a successful fighter must amalgamate two or more traditional martial arts categories to be successful. As there is currently no accurate way to describe a fighters style or understanding of the success of a style the first research question is:
- Can infight statistics and a fighters physical attributes be used to define styles and what is the difference in win rate of each style?
Additionally if styles are considered important predictors for a fight, differences between the attributes used to define styles are likely to provide important information in to what makes a fighter successful, prompting the second research question:
- What are the important physical attribute and skill differences between fighters when predicting the outcome of a fight?
Modeling method of winning as a crude approximation of style by weight class highlighted that weight was likely a factor in determining style. Higher weight divisions showed an increased likelihood fights would be won by knockout early and decreased the likelihood a fight would go to a decision. x x x
Conceptually this made sense as the more weight behind a punch the harder the hit.Examining physical attribute differences showed that older fighters were less likely to win a fight, with no observable difference in height within a weight class between xwinners and losers.
Infight statistics were highly correlated with one another showing the first hint of skills, i.e a fighter that threw a large number of punches to the head also through a large number of punches to the body, indicating he/she was a skilled striker. Clustering algorithms were applied to low correlation features and evaluated over the 3 main principle components.
K-means was unable to resolve clusters effectively whilst DBSCAN showed better resolutions with a small number of unclassified points, the averages of each cluster from DBSCAN were plotted on a parallel coordinates map, figure 5 and 4 distinct fighting styles defined in Figure 6.
Interestingly the points that DBSCAN was not able to classify were also added to the parallel coordinates with the next two highest win averages but outperformed them by 15%, as well as showing significantly better infight variables. Although thisbehaviour seems bizarre DBSCAN is noted in literature for it anomaly detection ability[6]. When evaluating fighters in this smaller cluster it appeared to contain the majority of the UFC champions.
Finally plotting the win rate of all the clusters (and the noise) to answer question 1 showed no significant difference in styles win rate with the exception of the outliers cluster, which had a significantly higher performance and was thus defined as a ‘Champion’ Cluster.
