Development and evaluation of a synthetic AI model for generating high-fidelity synthetic datasets

As the demand for data-driven insights grows across domains, there is a need for privacy-preserving data solutions. To address this need, the rise of synthetic data generation as a promising approach to creating datasets that retain the statistical properties of true data without losing privacy. In this research we develop and evaluate a novel synthetic AI model called the hangtime Activity generator, which generates high-fidelity synthetic datasets. Trained the generator with a broader, original dataset that includes the features of timestamps, accelerometer readings, and activity labels crucial for sports analytics and human activity recognition applications. This study aimed to further examine how much this generator can fit true statistical properties and complex relationships from the original data. Finally, our "hangtime Activity" generator produced an impressive overall accuracy of 97.6% and univariate and bivariate accuracies of 98.9% and 96.2%, respectively. These statistics demonstrate the ability of the model to maintain the integrity of each feature and its interactions. Performed a series of analyses to evaluate the fidelity of the synthetic data, including statistical summaries, distribution comparisons, and correlation analyses. The results showed high similarity between the synthetic and original datasets and little deviation in major statistical features. For clarity and ease of the analysis of the new dataset generated by clustering genes, distribution plots, and correlation heatmaps were used to explain the closeness of the synthetic data with the original dataset. The synthetically generated data set was analyzed using dimensionality reduction techniques, including Principal Component Analysis (PCA), and the data structure was visualized to show that the synthetic data efficiently captured the underlying patterns of the original data set. Moreover, the generator preserved the distribution of rare categories, such that these important but sparse data points remained well represented. Despite these promising results, limitations exist — such as the treatment of extreme values and, as always, the need for independent validation on more varied datasets. Future work will center around extending the flexibility of the model and considering the utility of the model in diverse domains. In general, our study advances the body of knowledge in synthetic data generation by building high-fidelity synthetic datasets for a broad range of data-driven applications and protecting data privacy and security.