Choosing a logistic regression trainer

Given the four logistic regression trainers available in ML.NET, which is the best for your problem? Whilst all four regression trainers return a binary classification, they are optimized for different datasets and workloads.

Are you looking to train and predict in a low memory environment? If so, the L-BFGS logistic regression trainer (LbfgsLogisticRegressionBinaryTrainer) is a logical choice given that it was created to handle memory-restricted environments.

Both of the SDCA-based trainers—SdcaLogisticRegressionBinaryTrainer and SdcaNonCalibratedBinaryTrainer—have been optimized for scalability in training. If your training set is large and you are looking for binary classification, either of the SDCA trainers would be a good choice.

The SymbolicSgdLogisticRegressionBinaryTrainer model is different from the other three in that it is based on a stochastic gradient descent algorithm. This means rather than looking to maximize the error function, the algorithm looks to minimize the error function.