Short Question About Implementation

[adam2392]

Hi @MrAE, @jbrowne6 and @falkben

Just pinging the ppl that seemed to touch these specific LOC.

I know you guys don't maintain this code anymore and have moved on, but I had a quick question in terms of what a specific line is doing. I was wondering if you could provide a quick answer (if you happened to write this part) to make sure I'm interpreting correctly. FYI: I have ported the code to cython and once this issue is resolved, I think we can safely move on :)

In

SPORF/packedForest/src/forestTypes/binnedTree/processingNodeBin.h

Lines 99 to 113 in a7a3c7e

	inline void randMatTernary(std::vector<weightedFeature>& featuresToTry){
	int rndMtry;
	int rndFeature;
	int rndWeight;
	int mtryDensity = (int)((double)fpSingleton::getSingleton().returnMtry() * fpSingleton::getSingleton().returnMtryMult());
	for (int i = 0; i < mtryDensity; ++i){
	rndMtry = randNum->gen(fpSingleton::getSingleton().returnMtry());
	rndFeature = randNum->gen(fpSingleton::getSingleton().returnNumFeatures());
	featuresToTry[rndMtry].returnFeatures().push_back(rndFeature);
	rndWeight = (randNum->gen(2)%2) ? 1 : -1;
	assert(rndWeight==1 || rndWeight==-1);
	featuresToTry[rndMtry].returnWeights().push_back(rndWeight);
	}
	}

are you sampling without replacement the feature index? It looks like rndFeature = randNum->gen(fpSingleton::getSingleton().returnNumFeatures()); can generate a random feature index, but is it possible to have a duplicate?

For example, say you have data with 4 columns, then maybe SPORF will sample a projection of:

indices = [0, 2, 0]
weights = [1, -1, 1]

Note that this in turn isn't a sparse linear combination with only +/- 1's, but now has a +2, -1 weight when doing the linear combination. Or is this function guaranteed to not have duplicates in its sampling of the projection matrix?

[MrAE]

Hey @adam2392, I worked mainly in the R part of things although I do remember having a similar issue with this chunk (lots of whiteboarding). I never did figure out if this block was sampling in accordance with the SPORF paper -- and given your example, I'd say it's not.

In that case the indices should be sampled without replacement -- going from memory.

I did tinker around in the C++ code, but the base functions came from James.
I know James had some code in his own repo, which may have some tests in it 🤷🏼‍♂️ -- he'd be the one with the most knowledge about how it works.