# meeting 22/12/2021 Overview: Multi arm bandit prblem Group items: similar to multi arm bandit problem boltzmann exploration -> exploration strategy generalize context example: type of user * given info arm does not work * though need prior info about user even if user is part of cluster of users -> need to explore but context helps explore to generalize to other users is it a business trip -> yes/no filtering flag: tells you the context amount of data is not the problem variable says something about: * how much someone is spending * could be per every euro in a bucket * might just be a few * could be less than and more than 100 euro * recommendations will be less targeted if its RS it's fine Try recsim * also dynamic, sequential sequential aspect of the recommendation recommend, watch, upate parameters, recommend as you interact over time, you get a better idea we could find coefficients recommend one time Direction: get something, start playing aorund with it and get an idea PhD start: will be easier to discuss with them how to transfer things between contexts all users are different, but we want to help the exploration from other user profiles collaborative filtering as basic techniques, look into othe rtechniques how to deal with big action spaces, you can show lots to people * score every movie, show top 10 * narrow down on what profiles they are balance certainty with potential fail for IG someone wants to spend lot of money, but we miss out on it by showing inexpensive hotels upper confidence bound takes max from this optimum ... certainty * if we don't konw what's going on, we explore * if we can select k of them * some function over movie space * if we keep sticking to one genre of movies, the user won't see other movies ever not our job to find the categories we want some simple categories **transfer across contexts** different ranking algorithms interleave, and compare bassed on what user chose **focus on sequential** context of the user, update the coefficients of the context based on the user if you explore too much don't build everything perfectly, that makes it hard to get ## Tasks > Get something, start playing around with it and get an idea 1. RecSim 2. LensKit Start reading documentation and apply them? or freeflow focus on sequential transfer across contexts