1

I am attempting to train an autoencoder on data that is extremely sparse. Each datapoint is only zeros and ones and contains ~3% 1s. Being that the data is mostly zero the autoencoder learns to guess zero every time. Is there a way to prevent this from happening? To give context this is extremely sparse data when you consider that the number of features is over 865,000

PDPDPDPD
  • 113
  • 2
  • 1
    while this is not a solution to your question, but a comment. Look for autoencoders used in building recommender systems. Recommender systems often use very sparse data (99.9% sparsity) as a tiny portion of the movies offered by netflix are of interest to a particular person. Neflix enginnering blogs could be a possible place. – user62198 Oct 15 '20 at 18:18

0 Answers0