Future photometric supernova surveys will produce vastly more candidates than can be followed up spectroscopically, highlighting the need for effective classification methods based on light curves alone. Here we introduce boosting and kernel density estimation techniques which have minimal astrophysical input, and compare their performance on 20 000 simulated Dark Energy Survey light curves. We demonstrate that these methods perform very well provided a representative sample of the full population is used for training. Interestingly, we find that they do not require the redshift of the host galaxy or candidate supernova. However, training on the types of spectroscopic subsamples currently produced by supernova surveys leads to poor performance due to the resulting bias in training, and we recommend that special attention be given to the creation of representative training samples. We show that given a typical non-representative training sample, S, one can expect to pull out a representative subsample of about 10 per cent of the size of S, which is large enough to outperform the methods trained on all of S.