Each directory contains a separate sampling procedure.
The full algorithm:
- In each layer:
- Compute m_k
- Draw m_k coalitions
- compute the MC in each coalition
- average the MC
- sum up the layers
The Monte Carlo sampling algorithm as presented by Ghorbani et al. in Data Shapley, adapted:
- Draw a permutation:
- Use all datapoints before i as the coalition
- compute Shapley Value for that coalition
- loop until convergence
A combination of Monte Carlo sampling with the standard algorithm.
- Compute the m_ks
- For each evaluated datapoint i:
- In each iteration:
- Draw a k with probability p
- Draw a sample S from k with uniform probability
- Compute the marginal contribution of i on S
- Loop until convergence
- Compute m_ks
- In each layer, sample until convergence is reached or m_k coalitions have been sampled
- move on