README to be perfected soon :)
The data are head-to-head results of international matches between the World Cup 2018 teams. The models we use are Gaussian priors and
- Logistic Regression with logit link likelihood
- Poisson Regression with log link likelihood
- Poisson Regression with soft-ReLU link likelihood
Our approximation techniques are
- Laplace approximation
- Gaussian variational approximation
- Metropolis algorithm
To predict, we use sampling methods to calculate the integrals - crude Monte Carlo integration.
We use PyTorch for its efficient tensor operations, but for purpose of education we will not use the efficient backpropogation methods.
wcAnalysis.py contains all the functions needed for all 3 approximations and its corresponding sampling functions
test.py runs the algorithm and runs the analysis
worldcup.ipynb was my development platform when I tested out my analysis. All the functions have now been properly implemented in wcAnalysis.py