Welcome back.

In the first video of this module, we introduced the concept of learning

recommenders, where we directly optimized a recommenders' parameters

to minimize its error or maximize its utility.

And there we talked about how not only can you optimize a recommender to have

the lowest prediction error or to do the most accurate job classification but

we can also optimize the recommender for its ranking capabilities.

We can directly optimize for

putting good stuff at the top of the recommendation list.

What we also said that there are some complicated subtleties to it and

we'll talk more about that later.

This is later and to help us with this we have Daniel Cloover who is a senior

PhD student in the Group Plans Laboratory and is also taking the lead

on lens kits implementation of Bayesian Personalized Ranking.

Daniel, welcome to the course.

>> Thank you Michael.

We're going to talk about Bayesian Personalized Ranking or BPR and the goal

of this is to build a personalized ranking function for each user.

And the reason we're talking about this algorithm in particular is it's

the most popular Learning to rank method.

That's a learning style recommender that uses ranking metrics,

instead of a prediction base metric.