Link Search Menu Expand Document
Recsys

Main Data Structure

These are the data structure highlighted in the SCRUF-D FAcct Paper, along with possible ways to implementing each structure in OBP.

User Recommendation List (l)

Machine learning algorithm (e. g. IPWLearner)

  • Context: User Profile (omega)
  • Output: Actions: itemIDs (v) and Positions: 0… N-1

Allocation History(H)

Bandit feedback collection of agent allocations with time index

Choice History(L)

  • Bandit feedback collection of user recommendation list (l) with time index
  • Bandit feedback collection of agent recommendation lists (l_f) with time index
  • Bandit feedback collection of choice function output list (l_c) with time index

Fairness metric for agent (m_i)

Off-Policy Estimator (OPE)

  • Context: Allocation History (H)
  • Context: Choice History(L)
  • Output: rating within [0,1]

Compatibility metric for agent (c_i)

Off-Policy Estimator (OPE)

  • Context: User Profile (omega)
  • Output: rating within [0,1]

Fairness Agent Recommendation Function (R)

Machine learning algorithm (e. g. IPWLearner) or hardcoded algorithm

  • Context: User Profile (omega)
  • Context: ItemID (v)
  • Output: rating

Allocation Mechanism

Machine learning algorithm(s) (e. g. IPWLearner)

  • Context: Fairness metric evaluations (m_F)
  • Context: Compatibility metric evaluations(c_F)
  • Actions: Fairness Agents (f)
  • Output: Agent allocation (beta) as action_distribution