An interesting area of online learning is a classification setting where the true label is not known and the algorithm must learn from partial feedback that is available. The Banditron applies ideas from the full information setting to the bandit model and balances exploration and exploitation. We plan to extend the Banditron algorithm which learns multi-class classification in an online supervised learning setting with partial feedback. The problem of particular interest to us is to predict r labels (using label rankings based on our model) and explore with a probability γ by replacing one (or maybe more) labels with a random label ranked lower by our model. This can be applied to a lot of real world applications such as recommendation systems where a user is recommended r relevant items with some random items included.

#################### Requirements ####################

1. Mongo (To avoid I/O reads from local file system, we are using MongoDB to store data)

2. PyMongo