For those of you who don’t know, Stumbleupon died about half a year ago. It was one of my favorite Internet “tools” where they send you to random websites on the Internet and you rate them thumbs up or thumbs down and the system supposedly “learns” your preferences over time and gives you websites more catered to your preferences the more you rate sites. And you can submit your own sites to the database as well which is how they expand the pool of websites to recommend to others.
This sounds like a Collaborative Filtering problem in a nutshell. I’m not sure how advanced Stumbleupon’s learning algorithm was but I highly doubt it was as advanced as Fast.ai’s collaborative filtering (after all, Stumbleupon launched over a decade ago.) And I see a lot of people all over the web complaining about Mix.com, the successor to Stumbleupon as being nowhere any good and more like a Pinterest clone.
Right now, I’m thinking of building a prototype for the backend to test and if it works, maybe hosting it as a simple web app. Does anyone have any recommendations for hosting services that will work well for this type of an app since it can potentially receive thousands or millions of URLs as user input to store in its database?
Edit: another question -
What’s the best practice for training on the additional user input that comes in over time from users rating URLs? Should I immediately train them into the model (with a batch size of 1) or should I batch them together in a queue and train the model periodically like when the queue exceeds a certain size?