Extending first paper

[stevehadd]

The part of the classification work to see what effect some small changes to the methodology of the first ML paper (Leahy & Llopis) makes to results. Examples of things to try for comparison

• training on a decade of data, or the whole time window of the dataset
• does the inclusion of the date have much effect on the performance when training one year at a time, compared to when training with longer time windows? Given the approximate nature of the year based threshold used in the iMeta algorithm, I would not expect that the exact date would be able to discriminate much within a year, but relevant over a decade or more worth of data.
• use scikit learn hyperparameter tuning to try different hyperparameters
• ensure class support is similar in training and test
• additional input features: quality flag, institute, platform, cruise

[stevehadd]

I'm happy that we have reproduced the first paper, and extended it through use of different splits and training on the whole dataset (rather than per year) with equivalent results. Further work with neural netwroks will be covered in other tickets, such as

• 27 additional features

• 24 cross validation and hyper parameter tuning MLP classifier

• 10 ANNs in pytorch