1
$\begingroup$

What sort of models do they use? Presumably some flavor of neural net. Do they do a lot of feature engineering? Or do they throw in huge raw matrices of oscilloscope output? Given that it must be supervised learning, what is their target? A thumbs up? How do they deal with heterogeneity in the population? We're not independent nor are we identically distributed.

It seems like it works better lately. Curious how it works.

$\endgroup$
1
$\begingroup$

Deep neural networks, probably with tens of internal layers, distributed across many many machines. Each user probably receives scores from many deep networks for various applications, and each network is probably automatically scored to determine when its particular model is most helpful. This is the pattern used across the industry. Music companies have the additional complexity of trying to parse sound data. Spotify does this with their Release Radar recommender.

As far as targets go, they probably build models to predict all forms of engagement. You probably get scores for listens, scores for time listening, scores for thumbs up or down, scores for add-to-list, and so on. Ensemble methods work well in practice and Google probably gives you on the order of dozens to hundreds of different machine-learned predictions.

| improve this answer | |
$\endgroup$
  • $\begingroup$ You say that industry standard is to score when a particular model is most helpful. Does that mean that each user is essentially given a weight, and google is using deep neural nets as sub-models in an EM algorithm? If that's what you mean, that's a lot of compute. $\endgroup$ – generic_user Nov 6 '17 at 22:20
  • $\begingroup$ Also how does Spotify's "Release Radar" actually parse the sound data? $\endgroup$ – generic_user Nov 6 '17 at 22:22
  • $\begingroup$ Check out nada.kth.se/~ann/exjobb/oktay_bahceci.pdf, which is "Deep Neural Networks for Context Aware Personalized Music Recommendation" by Oktay Bahceci. In short, Spotify uses the sound file as input to a neural network. Here's more information on WaveNet, an audio neural net from DeepMind: deepmind.com/blog/wavenet-generative-model-raw-audio. $\endgroup$ – Dan Jarratt Nov 6 '17 at 22:38

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.