AutoBNN: Probabilistic time series forecasting with compositional bayesian neural networks

Researchers have introduced AutoBNN, a powerful and flexible framework that combines Bayesian Neural Networks (BNNs) and Gaussian Processes (GPs) to build advanced time series prediction models. This innovative library, developed by Colin Carroll, Thomas Colthurst, Urs Köster, and Srinivas Vasudevan, leverages compositional kernels to enable complex data modeling and forecasting. AutoBNN simplifies the process of creating sophisticated models by offering a scikit-learn-inspired interface. Users can easily define models using operators such as Add, which allows for the combination of different BNNs like Periodic, Linear, and Matern BNNs. The framework also supports both Maximum A Posteriori (MAP) and Markov Chain Monte Carlo (MCMC) inference methods, making it adaptable for various data types and requirements. One of the key features of AutoBNN is its use of WeightedSums, which allows for the exploration of rich combinatorial structures without significantly increasing the number of learnable parameters. For instance, the 'sum_of_products' model enables the creation of up to 2^16 different discrete structures by turning base kernels on or off, all within a single model. The library includes support for multiple likelihood functions, including normal distributions with different noise characteristics and negative binomial distributions for count data. This makes it suitable for a wide range of applications, from continuous data forecasting to count-based predictions. AutoBNN has shown promising results on real-world datasets, such as the Mauna Loa CO2 dataset, where it effectively captured both trend and seasonal components. The framework's ability to provide uncertainty estimates makes it particularly valuable for decision-making in complex environments. The open-source library is available on GitHub and comes with a Colab example to help users get started quickly. The team behind AutoBNN invites the broader community to explore its capabilities and contribute to solving real-world challenges through innovative time series forecasting techniques.