time_id is null in Demographic covariates for temporal covariates about featureextraction

[schuemie]

Hi Chan!

Yes, that is intentional (the demographics typically do not pertain to a particular time window).

Would it be possible to modify CIReNN to work with the NULL values? (Maybe convert them to whatever CIReNN expects just prior to fitting?)

[jreps]

I can modify the way the data get mapped. I'm not sure how best to include non-temporal data into the CIReNN though? Should we have a non-temporal matrix (people by non-temporal features) in addition to the temporal tensor (people by time by temporal-features)?

[jreps]

Just spoke to Peter and he had a great idea. How about rather than putting NULL as the time id for demographics, having the demographics for each time id and for things like age, make that days and have that as the actual value on the time id setting. Does that make sense?

[chandryou]

@jreps , That's just what I've thought! If so, I don't have nothing more.

[schuemie]

@jreps , we can definitely do that, but it would mean a huge increase in size of the covariateData objects because demographics tend to not be sparse. For example, every subject will get an age for every period, and every subject will get a gender for every period.

If all temporal algorithms require this, then we should do it at construction of the features. If only a few do, we should probably write a function that expands the current compact representation into this new large format only for those algorithms that need it.

[jreps]

I was thinking, I can do it when I convert the data into the array, I'll update the plp code and test next week