The application of machine learning and deep learning in sport: predicting NBA players’ performance and popularity

Figures & data

Figure 1. Distributions of cross-validation results for train data from six candidate regression models.

Figure 2. MAE and RMSE results for valid datasets from candidate regression models.

Figure 3. Top 10 Relative features for final regression model by traditional machine learning.

Table 1. Cross-validation results on random-, SMOTE over-sampling and under-sampling without predictor variable ‘being selected for this season all-star game’ for train data.

Models	Random over-sampling				SMOTE over-sampling
	Precision	Recall	ROC AUC	F1	Precision	Recall	ROC AUC	F1
LG	0.892	0.920	0.900	0.910	0.900	0.936	0.918	0.916
SGD	0.890	0.910	0.900	0.900	0.883	0.951	0.913	0.916
LSVM	0.880	0.930	0.904	0.910	0.890	0.948	0.916	0.918
PSVM	0.885	0.930	0.904	0.907	0.901	0.954	0.924	0.927
RF	0.980	1.000	0.9900	0.9900	0.959	0.981	0.964	0.970
NB	0.806	0.910	0.845	0.855	0.810	0.921	0.852	0.862
AB	0.905	0.918	0.911	0.911	0.910	0.932	0.920	0.921
GBM	0.918	0.970	0.9420	0.943	0.926	0.959	0.941	0.942

Table 2. Cross-validation results on random-, SMOTE over-sampling and under-sampling with predictor variable ‘being selected for this season all-star game’ for train data.

Models	Random over-sampling				SMOTE over-sampling
	Precision	Recall	ROC AUC	F1	Precision	Recall	ROC AUC	F1
LG	0.8893	0.8870	0.8883	0.8881	0.8912	0.8985	0.8944	0.8948
SGD	0.8503	0.9376	0.8862	0.8918	0.8608	0.9463	0.8967	0.9016
LSVM	0.8818	0.9018	0.8904	0.8917	0.8829	0.9122	0.8956	0.8973
PSVM	0.8821	0.9034	0.8913	0.8926	0.9009	0.9538	0.9266	0.9244
RF	0.9807	1.000	0.9901	0.9902	0.9560	0.9910	0.9727	0.9732
NB	0.8065	0.8932	0.8395	0.8476	0.8066	0.8998	0.8420	0.8507
AB	0.9054	0.9075	0.9064	0.9065	0.9059	0.9153	0.9106	0.9101
GBM	0.9163	0.9670	0.9394	0.9410	0.9225	0.9544	0.9371	0.9382

Table 3. Results on valid data from candidate classification models without predictor variable ‘being selected for this season all-star game’.

Models	Precision	Recall	ROC AUC	F1
RF	0.4881	0.6308	0.7953	0.5503
GBM	0.3518	0.8308	0.8689	0.4943
BB	0.3652	0.8231	0.8680	0.5059
BRF	0.2977	0.9000	0.8855	0.4474
RU	0.2418	0.2846	0.6152	0.2615

Table 4. Results on valid data from candidate classification models with predictor variable ‘being selected for this season all-star game’.

Models	Valid data
	Precision	Recall	ROC AUC	F1
RF	0.5525	0.7042	0.8343	0.6192
GBM	0.3927	0.8380	0.8785	0.5348
BB	0.3980	0.8521	0.8858	0.8426
BRF	0.3155	0.9155	0.8957	0.4693
RU	0.3524	0.7817	0.8459	0.4858

Figure 4. Top 10 Relative features for final classification model by traditional machine learning.

Figure 5. Loss function results for train and valid data and evaluation loss for test data from each candidate deep learning model.

Figure 6. ROC curve for train data from each candidate deep learning model.

Figure 7. Loss function results for train and valid data and confusion matrix results for test data from each candidate deep learning model.