> I recommend you start with the basics (datasets, features, training/test/validation data splits) and don't worry too much about the actual choice of model---t

> I recommend you start with the basics (datasets, features, training/test/validation data splits) and don't worry too much about the actual choice of model---there will always be shiny new models with better performance but sometimes using the "old stuff" is good enough. Or, conversely, it may be that all models are just as bad. This seems to be the case in my domain (formal proofs), where the bottleneck seems to be data representation; it doesn't matter which learning algorithm you use, when your feature selection has stripped out all of the learnable information ;)