I will present our work on highly-scalable out-of-core techniques for learning well-calibrated Bayesian network classifiers. Our techniques are based on a novel hybrid generative and discriminative learning paradigm. These algorithms
- provide straightforward mechanisms for managing the bias-variance trade-off
- have training time that is linear with respect to training set size,
- require as few as one and at most four passes through the training data,
- allow for incremental learning,
- are embarrassingly parallelisable,
- support anytime classification,
- provide direct well-calibrated prediction of class probabilities,
- can learn using arbitrary loss functions,
- support direct handling of missing values, and
- exhibit robustness to noise in the training data.
Despite their computationally efficiency, the new algorithms deliver classification accuracy that is competitive with state-of-the-art in-core discriminative learning techniques.