Evalml: Add metrics: RMSLE/RMSE/MSLE

[gsheni]

Add:

• root mean squared error
• root mean squared log error (which happens to be the default xgboost metric)
• mean squared log error

[dsherry]

@ctduffy raised a good point in slack this afternoon: any metric which first takes the logarithm (MSLE/RMSLE) is only valid when both the target and the predictions are guaranteed to be non-negative. Note that sklearn's impl will throw ValueError: Mean Squared Logarithmic Error cannot be used when targets contain negative values when its provided with input containing a negative value.

I propose the following:

• Implement MSLE and RMSLE objectives, as subclasses of RegressionObjective. They can both use sklearn's mean_squared_log_error, so they'll throw an exception for negative input.
• Make sure the documentation for MSLE and RMSLE explains that those metrics are only well-defined for non-negative input.
• Update AutoRegressionSearch.search to throw an exception if the objective is MSLE/RMSLE and the input y target contains any negative values
• Make sure AutoSearchBase._evaluate gives a pipeline a nan score if the objective is MSLE/RMSLE and any of the model predictions are negative
  
  
  
  • I think no code changes will be required for this, as long as the objectives throw an exception. Why: AutoSearchBase._evaluate tries to use the objective to score the pipeline. If exceptions are generated, the pipeline gets a nan objective score.
  
  

@ctduffy @gsheni does this sound good to you?

Also, this is exactly the sort of check which could use our new DataCheck API in the near future! @angela97lin is still working on #709 which adds the data checks to automl, so rather than dig into that now, my recommendation is that we file a separate issue to track porting the proposal above to use a DataCheck once #709 is done. @angela97lin , would you mind filing that?

[angela97lin]

Filed #785! I think what @dsherry proposed is good--if we include good documentation stating that valid inputs must be non-negative then it's very reasonable to throw a ValueError!