A great move for Google would be to treat notifications in Android as Google Now cards; i.e have some machine learn if it is relevant content. Observing if the

A great move for Google would be to treat notifications in Android as Google Now cards; i.e have some machine learn if it is relevant content. Observing if the user has opened the notification, and also offer more feedback options in the long press menu. IMHO, best results would be obtained by letting the apps emit their own feature vectors to be fed to the learner. Google could aggregate these among all their users. App creators would learn to provide features that actually inform of a kind of proposed interaction. Failing to do so would boot them out of this users' "attention market". I receive about 20 useless facebook notification for a relevant one (mainly when someone is talking to me in the fb chat). Facebook could clearly publish the following along to the learner with each notification: * is it a chat message? * is it from someone I have talked to recently? * is it from a cluster of users that talks to each other frequently that I am part of * a representation of the cited cluster above * location of the content (with some additional feature engineering from the phone, who knows if I've been there previously) * it is a photo/video/text? * a doc2vec vector if it is text content * an image embedding if it is a photo etc. I believe this could reduce the amount of notifications by a solid order of magnitude, while not lowering the usefulness value too much. We would suffer way less hijacking, and reduce our addiction of the mindless novelty-seeking kind.