Feature Idea: log-link models about robyn

[gufengzhou]

Hi there, you're totally right about the log-linear model, which can be specified in glmnet as family = "poisson". The reason why we haven't touched this part is that first we lacked real count data for the dev at the beginning and second how to do decomp is still quite controversial. Would you elaborate your approach to decomp a multiplicative equation?

[lorenze3]

I've used different approaches. One that is surely not proprietary is something like the below, written in psuedo code for clarity, happy to re-write if not clear.

yhat=predict(X)
for thisX in X:
newX=X
newX[,thisX] = 0
thisXDecomp_step0=yhat-predict(newX)
Decomps_step0[,thisX]=thisXDecomp_step0
end

now Decomps_step0 has all step0 decomps in it, but of course the total of those may not equal yhat
Decomps_step0[,psuedoPred]=sum_across_all_step0_columns
Decomps_step1=Decomps_step0
Decomps_step1[,adjFactor]=Decomps[,psuedoPred]/yhat elementwise division, here-- adjustment factor to get total to equal original prediction

_for each _step0 column in Decomps_step1, make a step_1 column = step0*adjFactor

for step0 in Decomps_Step0:
make step1 name
Decomps_step1[,step1]=Decomps[,step0]*Decomps[,adjFactor]
end

[gufengzhou]

Thanks for sharing your experience! Providing multiplicative model is on own roadmap, as well as other interesting options like Bayesian model for better calibration, state space model for time varying coefficients etc. For the moment, an update feature to help refreshing the model regularly will be coming next.