# Meeting 2

## we don't say we were busy
## Focus on explainable recommenders
* Compare existing ERs
* Dive into knowledge graphs

survey: Explainable Recommendation: A
Survey and New Perspectives
1. Explainable Deep Learning for Recommendation
2. Knowledge-enhanced Explainable Recommendation
3. Multi-Modality and Heterogenous Information Modeling
* using heterogenous data not just for recommendation but for explanations as well
4. Context-aware Explanations
* just like how the recommendations themselves are context aware (for example using time and location), explanations could be context aware as well.
* current explainable recommendation models are static (users are profiled on a training set and explanations are generated based on that)
* It could be based on exploration as well
5. Aggregation of Different Explanations
* Different explainable models for different explanations
* Challange: select best combinations of explanations for a recommendation
* Integrate statistical and logical reasoning approaches into machine learning
6. Explainable Recommendation as Reasoning

## how is phd recruitiment going

## something concrete to submit now
* RLRS restrained by the fact that it has to be explained
* evaluation accuracy should be metric of RLRS -> dude i fucking love the way you phrased it
* trade-off
* multi-objective reinforcement method, find the
* optimize explainability and value of recommendation
* not easy to map onto one number, thus a *multi-objective* approach
* take a look at multi-objective RLRS
*
## PHD stuff
* parallel stuff, one agent with multiple environments
* [CONFUSIION]
* generalisation/ mapping
* trying to do exploration better
* explainability isn't really on there
* multiple types of users
* find out what they are -> supervised learning
* **figure out which one you are interacting with -> quicker you find out, better recommendations (cold start)**
* RL: trying to learn the way to distinguish between these users
* speeding up the cold start
## Ideas
* study on exploration for recommender systems

Reinforecement learning and recommmender system, interesting: doing many things in parallel, interacting with multiple users at once. Can't look at all combinations: too much data.
Who to do the experiment on?
Sketch broad context
Recsim as a platform, so they see the first steps.
bunch of research questions with first steps.

## stuff for meeting

model-based exploration
exploration vs exploitation in the context of multiple environments
coldstart
different complexity in training environments
recommender for multiple environments
* connect starting conditions to the final generalizability of the recommender systems, find out what specifically caused it, in other words connect end result with different design choices we made at the start. We were inspired by this article, in which one of the conclusions was that the more complex the starting environment was for the agent, the more generalizable it became on later unseen tasks.
* two agents: one starts with 0 data, the other with accounts
* multi-environment: (multiple users/ same agent or model)
* speeding up the cold start, finding what group the user belongs in
* exploring exploration: multi-objective RL: take into account how much information a specific recommendation gets you
* evolutionary RL (need to find parallels in the evolutionary tasks and rewards to recommender tasks and rewards)

Our research will be on reinforcement learning based recommender systems. More specifically, we are going to explore the connections between starting conditions of the agent and the final generalizability of its policy to different users. The starting conditions could vary in terms of access to different types of data, and different reward functions. We would like to produce new findings about how different kinds of design choices made at the start lead to different end results in multi-environment scenarios.

// *and how they can be tweaked to alter the end result depending on the needs of the recommender system.* //
*We were inspired by this article, in which one of the conclusions was that the more complex the starting environment was for the agent, the more generalizable it became on later unseen tasks.
We are aware that the proposal above is quite general, but like we discussed in the meeting, we are planning on refining as we progress in our research.*