useR! 2024

Generative machine learning has gained world-wide attention and, especially since the rise of ChatGPT and DALL-E, has started to become an integral tool both in business and everyday life. While the hype has mainly focused on text, image, audio and video synthesis so far, generative modelling of mixed tabular data with both continuous and categorical variables has great unexploited potential in many research fields and industry applications. However, recent attempts to adapt the existing, mainly deep learning-based methods to this more general setting have not shown the same overwhelming successes yet. We present the CRAN package ‘arf’, an easy-to-use implementation of adversarial random forests based on ‘ranger’, which has shown the ability to match and often outperform current deep learning approaches in terms of performance, tuning efforts and runtime, also on small or high dimensional data. ‘arf’ provides tools for both synthetic data generation and density estimation. Optional conditioning on events further extends the possible area of application, enabling for use cases like missing data imputation, data balancing and augmentation.