What machine learning algorithm should I use to build an intelligent reminder? [closed]

If you're sure you want to go with machine learning for this, it's only going to work (in the sense that users care about) if the user can give precise and immediate feedback about the usefulness of each notification.

In particular, when the user receives a notification, ideally, they should be able to immediately click "This was useful" or "This was not useful"; the latter may provide a few optional reasons to choose from like "too early", "too late", "no reminder needed", etc. And they should always have the option of overriding the machine learning with manual configuration if they feel the need to ("anything with 'dentist' should have a reminder 1 hour in advance, period").

The main problem is that on many mobile devices, you have to unlock the device before you can respond to the notification in any meaningful way, which means users aren't going to give you that feedback a lot of the time. For that reason alone, I think you should probably focus on manual configuration.

If you're talking about a user training an AI to learn what's important and what isn't based off of their calendar and some sort of feedback, then it's going to be a lot less work for them to just mark the appointment as important/not important with a flag or something when they're making it.

If you're talking about an AI trying to learn what is and isn't important based off a single calendar without any other input, you don't have enough data to learn anything.

It still isn't feasible even if you have it learn on EVERYONE's calendar without some sort of feedback. You could hand-pick a few examples you thought were important and have the system learn what you think is important, but that's awfully slow.

The most viable path I can see for meaningful machine learning is if you had an program have access to everyone's calandar and at least some of the people were anal about marking their appointments as important or not. The thing would learn based on the examples provided how to classify events as important or not.

And you've then created a spam filter for calendar events.

Check out: http://en.wikipedia.org/wiki/Naive_Bayes_spam_filtering for a method to pick out important or unimportant things. It's all about keywords. "Doctor", "Appointment", or "MUST GO" would be highlighted while "beer run", or "bill due" would be muted.

Machine learning is pretty complex though. It'd be simpler to provide the users some control of some rulesets for what appointments get reminder for when. After you've got something usable, and users entering in rulesets you can harvest, then you might have something you can learn from.

Personally, I'd like to see such an app make some guesses about WHERE the appointment is located along with my current location and distance to the target. If I'm a hour away from the dentist, getting a reminder 15 minutes before I'm supposed to be there doesn't help much.

Good luck!