Mitigating Bias and Error in Machine Learning to Protect Sports Data

One of the essential processes in modern sports is doping control. In recent years, specialized methods of artificial intelligence and large-scale data analysis have been used to make faster and simpler detection of violations of international regulations on the use of banned substances. The smart systems in question depend directly on the quality of the data used, as high-quality data will produce algorithmic approaches of correspondingly high quality and accuracy. It is evident that there are many sources of errors in data collections and intentional algorithmic interventions that may result from cyber-attacks, so end-users of artificial intelligence technologies should be able to know the exact origins of data and analytical methods of these data at an algorithmic level. Given that artificial intelligence systems based on incomplete or discriminatory data can lead to inaccurate results that violate the fundamental rights of athletes, this paper presents an advanced model for mitigating bias and error in machine learning to protect sports data, using convolutional neural network (ConvNet) with high-precise class activation maps (HiPrCAM). It is an innovative neural network interpretability technique, wherewith the addition of Bellman reinforcement learning (BRL) and Broyden–Fletcher–Goldfarb–Shanno (BFGS) optimization; it can produce high-precision maps that deliver high definition, clarity, and the input and output capture when the algorithm makes a prediction. The evaluation of the proposed system uses the Shapley value solution from the cooperative game theory to provide algorithmic performance propositions for each of the produced results, assigning partial responsibility to parts of the architecture based on the impact that the efforts have on the relative success measurement, which it has been preset.