Explore effect of cruise ID feature on classification results

[stevehadd]

There seems to be a correlation between type and cruise ID. Cruise ID has not previously been included as a feature in classification. Its effect on results should be explored. How cruise data is split between training and test could be important. The following for be tried

• use only cruises where all points are labelled so we can see the results
• split all cruises between train and test (do all cruises have multiple profiles?)
• have some cruises only in train, some only in test, some split between train and test
• having some only in test is very important because some cruises have no model/manufacturer metadata. We want to know if our model is skilful useful for such cruises, and whether it is more skilful than the deterministic iMeta algorithm in these case in particular.

[stevehadd]

The ability to split on cruise ID has now been implemented in the dataset class. The performance of the ML algorithms on unseen cruises has been demonstrated for decision trees and neural networks. Further analysis of the relationship between cruise ID and performance will explored in other issues.