You can use correlation or the coefficients of a linear model iff the features on the same scale. Another method is to train a model leaving out each feature once, then you see how much accuracy drops.