# First meeting M

## Three topics
1- explainable recommenders
2- current top k recommenders
3- off-line policy evaulation techniques for top k recommenders

* What they are doing in mercury lab?
* Is there something that we can do to assist them there?

been a while since supervised. -> send him the link

RLRS
introductions, then tell him what we want, find PhD students
teaches Automata Computability and Complexity
AD: Smart algos for smart decision
RL, we need the specific algorithms
Making general algorithms

safe RL interactions with environment, avoid things goinog wrong
* more careful in exploration
* or train based on offline data
* real problems, getting data is hard or the amount of effort would be hard
* decision support system
* mercury lab: hiring PhD students, problems that booking.com is interested in.
* booking has a lot of data
* non-stationary
* Diverse but focused: Causality
* not classifying users (not our job), but using their data

Propensity score matching:
Student few years ago:
* you need the old policy, no exploration
* off-policy for exploration
* top-k stuff
* PhD students are still being hired, so we don't know who does anything, so let's wait for now
* This algorithm looks like one, but there's something missing
* Evaluation: problem: always interfering with the system, so you need a system that is decent enough (?)
* RS simulator
* problem we wanna solve, solution ?, steps to take, and test it using a simulator <- best way to go forward, meet a few weeks from now

https://github.com/google-research/recsim is it any good? does rs, and built by people who care about RL
* allows testing
**Not** building simulators on our own

Questions for next meeting
* Fail if didn't do the right methodology, not if nothing comes out of the research
* Milestones are nice, shows progression:
* 1.
* 2.

Good to have a plan.

* Applications in deep learning in mercury lab
* could look at modeling or algorithm part
* lot of magic there still
* maybe as an add-on: compare our project to DRL