search

SWS-Methodology / faoswsFlag

This is an add-on package to the faosws library to support the manipulation and the aggregation of flags.
0 stars 0 forks source link

Improve efficiency of flag aggregation #9

Open chrMongeau opened 7 years ago

chrMongeau commented 7 years ago

I like to share here an issue in the trade module, as I think it should be addressed at faoswsFlag level. The issues is:

https://github.com/SWS-Methodology/faoswsTrade/issues/154

As I said there, I think the aggregation should be re-worked by rethinking the flags to be ordered factors. I probably miss a lot of the functioning of faoswsFlag, so I don't know if it would clash with the package design/goals. Anyways, given that most of the code would break if the flags are expected to be factors, maybe there should be an option to aggregateObservationFlag() so that when it has as input ordered factors then it's behave in a predetermined way.

@sebastian-c what do you think?

sebastian-c commented 7 years ago

While an ordered factor makes some kind of intuitive sense, I'm not sure if it will always be that simple. Is there a reason that aggregateObservationFlag can't convert its arguments to ordered factors internally and then convert back to character at the end?

mkao006 commented 6 years ago

Thought I might jump in while I am here ;)

The reason why the weights were numeric rather than ordered factor was because they were used to construct the distribution of each value. These distributions can then be used to balance the SUA or the FBS.

However, nobody understood how it works before I left, so I guess it is no longer relevant. The details of the methodology can be found in the vignette. But you have to compile it yourself.

mkao006 commented 6 years ago

Here you go, had to Google how to compile an Rnw. This is like last century technology .....

faoswsFlag.pdf

See section 4 for the reason behind numeric values.

Here is the distribution paper.

chrMongeau commented 6 years ago

The reason why the weights were numeric rather than ordered factor was because they were used to construct the distribution of each value.

Thank you @mkao006. In this context, it makes sense. AFAIK, no one is indeed using flags in order to do what is suggested in the framework. Anyhow, for now I'm not touching anything in the design as the bottleneck in the original problem is probably somewhere else. As soon as I get time, I'll investigate further.