A performance evaluation of neural network features and functions settings on the model accuracy

Not only in sports is a neural network the most used type of artificial intelligence. With software development, anyone can create a neural network model, but little is known about how to prepare the data and how to set up the model algorithms to their maximum performance. For these reasons, this study aims to determine whether features or function settings have a greater effect on model accuracy. An initial feature dataset (n = 18882) was obtained from publicly available sources. Each of the six different feature settings consisted of 96 models. A total of 384 models were created, in which their testing accuracy and the percentage difference between the training and testing phases were further analyzed. No statistically significant differences were found between the accuracy of the function’s settings, but statistically significant differences were confirmed between the feature settings. The study found that feature settings, especially the reduction of the number of outputs, are a more important factor in increasing the model accuracy, than function settings. Although the literature focuses more on the function setting and sets feature setting is taken rather as a type of how to improve the model.


31
With the growing general public's interest in sports and technological development, a huge amount of 32 structured and unstructured data is emerging that can be used for predicting future events [1].

33
Although, artificial intelligence (AI) has been used for over 20 years is still considered a relatively 34 new (but rapidly evolving) technology capable of simulating human intelligence [2]. Machine learning 35 (ML) is an integral part of AI and refers to the automated detection of patterns in datasets [3]. ML was 36 first characterized as a field of study, that allows computers to learn without being explicitly 37 programmed [4]. The newer definition characterizes ML as a computer programming process to 38 optimize performance criteria using (mostly) data [5]. An increased interest in predicting sports 39 outcomes started around 2010 [1]. ML is typically classified as supervised learning, unsupervised 40 learning, and reinforcement learning. It is also possible to encounter an additional division into Semi-41 supervised learning, Transduction, and Learning to learn algorithms and more [5][6][7]. This study is a (employees or services of another company) for data analysis and so they were able to (successfully)

128
Training and testing accuracy (%) were obtained for all calculated models (n = 384). Along with the 129 results of the percentage difference (%diff) between training (V1) and testing (V2) accuracy. 130 The most accurate models (from the testing accuracy) were more thoroughly analyzed using The

134
The assumption for the parametric analysis of variance test was violated according to the results of  differences between the training and testing phases gradually increased, until the output setting (see 167 Fig 2), a model can be also described as overfitting [1,19]. Which was reduced by the 5 th and 6 th 168 phases, more precisely, by reducing the number of outputs. Therefore, the reduction of the possible 169 outputs could be described as a suitable technique not only to increase the accuracy of the model 170 prediction but also to reduce the chances of model overfitting.

171
172 Table 1. an overview of the most accurate models in the four individual phases of strengthening.
173 prediction decreased. This opposite conclusion may be because the model from this study has not yet 177 reached its greatest accuracy. Further in Fig 1 and 2 can be seen, that after removing the least 178 important features from the model, neither the accuracy nor the difference changed significantly.

179
According to the results of the descriptive analysis, it can be argued that the accuracy (of the testing 180 phase) is similar between raw data (x̅ = 7.2 ± 1.2; skewness = 0.1; kurtosis = 1.0) and data without can distinguish between outcomes (classes). Therefore these 41 most accurate models were further 208 evaluated using ROC Curve and AUC, which ranged from 0.125 to 1.000 (AUC). The following Table   209 2 provides an overview of models (n = 14) that achieved 100% accuracy and AUC was equal to 1.00.
210 211 Table 2. Overview of the model settings with 100% accuracy and 1.00 AUC.

213
It seems that the best models are composed of two hidden layers and Hyberbolic tangent as the was not found. Therefore, it maybe depends more on the data and the situation. Interestingly, Table 2 217 does not contain an Identity and Hyperbolic tangent output layer. would be updated weekly, should provide more relevant predictions.

258
The inability to compare the properties and accuracy of two different models (since they do not come 259 from the same dataset) significantly complicates the work of researchers and data engineers.

260
Therefore, it would be appropriate to invent or derive some methods, approaches, or strategies for this 261 type of analysis. For example, based on the effect size testing.

263
Due to the huge popularity of artificial intelligence and machine learning, the number of statistical