How can I create new funny pictures?

How can I create a pictures recommendation system?

• I have an app which provides wallpapers to users. They can choose the picture they like. To make the system more clever, the user can "like" the picture if they love this picture. So how can I create a pictures recommendation system? I only have the "like data". I plan to give tags to all the pictures. Then we will recommend the tags that the user likes. Am I right? Are there better ideas?

• Answer:

  This is a classic collaborative filtering problem. You have so-called "implicit" data, as likes, and not ratings. That's fine. You do not even need tags; the likes can be enough. Although I would strongly encourage you to use clicks as weak "likes". You will have much more click data and will mine a lot of value from it, and I suspect you will not have much like data at first. As to how to architect it, that is a very big question. At small scale, I might refer you to the parts of Apache Mahout that are not distributed -- the org.apache.mahout.cf.taste package, describe in Mahout in Action. This is an easy way to make a small embedded recommender in Java. If you mean to architect for large scale, my best answer is to point you to my follow-up project, Oryx. It is more complex, but exactly suited to your implicit data use case, and does not need Hadoop either. The architecture is discussed briefly too: https://github.com/cloudera/oryx