Perform hyperparameter tuning on L2G

As a developer, I want to run cross-validation and hyperparameter tuning on L2G because this will allow us to enhance the model performance.

Background

The L2G model is based on a gradient boosting (XGBoost) algorithm that typically benefits from careful tuning of hyperparameters to obtain an optimal performance. Our current set up allows us to parametrise variables like the depth of the tree, learning rate, or to apply regularisation values just by defining the parameter grid through the configuration file. At the same time, with W&B we will be able to evaluate how the model is reacting to the different parameter combination.

Personally, I don't expect these will play a huge role in the performance of L2G, but I hope to see subtle improvements in model accuracy and the ability to generalise.

Tasks

[ ] Run cross validation on a grid of hyperparameters. The model will be trained and tested for each possible combination of hyperparameters
[ ] Analyse the results in W&B to understand how specific variables impact the performance of the model
[ ] Update the L2G step configuration with the set of parameters yields the best results

Notes

The work describe in https://github.com/opentargets/issues/issues/3252 will have an impact on this issue. I'd suggest attacking that one first.

opentargets / issues