seasonality in the training dataset - convert month as a predictor column?

[JoeRibeiro]

A functional group may get larger, more or less fluorescent during a certain months. This interseasonal variability will be species specific and zone specific. But it affects where the cytogram boundaries lie. One way to train a better model would be to have the month as a predictor. This is risky though, as we start to mix the particle's physical characteristics with predictors that may cause odd behaviours like being more likely to predict a bloom species at any particle in the cytogram just because it is april.
...Perhaps we could have a different model trained for each month?
This one is a "further research" question.

[JoeRibeiro]

On reflection, I think this may be a limitation of using listmode parameters only. If we are using random forest, the forest may vote that the a particle is one class if variable x >0.5 and a different class once x < 0.5. Adding a month/time variable in to the prediction may work, but for the reasons above, it would not be robust, and it is preferable to use more of the available data associated with the particle to make this decision.

That brings us on to CNNs for (a) the images and (b) the pulse shapes. It would be preferable to use those data to inform the classification decision. It would be a more robust way of determining if the same species is present in the water but it is just fluorescing less this month than last month. Or if the sample contains a new species / functinoal group.

Including info from the pulse shapes and images in CNN would be a project in itself though.

[LucindaLanoy]

I thought of this towards the end of my Master's internship, except I had a straightforward "season" column. While calculating feature importance using RFE, I found that this column had some kind of high predictive value.

However, when I shared my results, my colleagues told me that having a seasonality predictor could be dangerous, as all the global changes mean that seasons (or months) might not be reliable enough. Instead, parameters that contribute to this seasonal effect could be incorporated, such as temperature, which was measured at the same time as the sampling.