Multi-task learning

[jaydu1]

• Test for multi-task learning in two scenarios:
  • [ ] a MultiOutputRegressor would learn an estimator for each target individually.
  • [ ] aRandomForestRegressor or DecisionTreeRegressor will be different: when defining a split at a tree node, the criterion computing will take into account all outputs by summing the criterion for each output. Therefore, the split is regularized since it will minimize the criterion on all outputs.
• Add different criteria for multi-task learning
  • [x] #3
  • [ ] individual MSE
• CV results
  • [ ] Summarized outputs/results for multi-task learning
  • [ ] Model refitting with best parameters

[jaydu1]

For individual MSE fitting, use a wrapper function as in random_forests.ipynb.