Link Search Menu Expand Document
Recsys

SCRUF-D Integration

Similarities

  • Both reference logged data: Both use feedback logs to improve predictions
  • Both use similar data structures: UserID, Lists

Differences

  • OBP uses reward system
  • OBP is based on Machine Learning driven by data and SCRUFF-D is more model based
  • Metrics (OPB is rewards based)

Possible Implementation

All the following are various ways of integrating OBP into the SCRUF-D architecture. (not currently implemented, but are possible solutions)

Partial

  1. OBP as recommender system
  2. Build Algorithm and Choice Mechanism, feedback separately

Full

  • Modifications (Similar to Slate)
    1. Evaluators/ Metrics
    2. Bandit Feedback
    3. Determine data sources/storage
  • Ways to manage the Rewards Problem
    1. Set all rewards equal to 1
    2. Repurpose the rewards label: Could signal protected items/lists
  • Ways to manage the Fairness Agents
    1. Each Fariness Agent serves as an OBP policy
    2. Allocation Mechanism manages policy usage based on BanditFeedback: Nested actions
  • Ways to manage the List Structure
    1. OBP uses three dimension (context,actions,posion): SCRUFF-D uses (userID, ItemID, pposition)
    2. Context,ListID could be used in OBP, but we would need to pull list from data source