Tolerence error;

[wangk28]

When fitting BetaGeoFitter with small sample size data, I always get the following message:

Desired error not necessarily achieved due to precision loss.

Thus resulting in ConvergenceError. When this error occurred, I have to manually set the a high tol in BetaGeoFitter and GammaGammaFitter. Is there a way that the tol can be automatically changed when facing precision problems?

[psygo]

With small sample size data, it's going to be more difficult to have the model converge. There probably isn't enough consistency in the data for the parameter fitting to find a stable combination. You can set a high tol, but you should actually be more concerned with getting more data instead.

Also, if you get some more data, try toying around with the penalizer_coef before messing with the tol.

[wangk28]

thanks for detailed explanation. However, the use case here is that there exists some extreme customers and behave quite different than others(highly-paid customers) which only covers a really small proportion of the training data(~0.1% and ~20 samples). Will high 'penalizer_coef' help solving such problem?

[psygo]

If these customers are such a low percentage of your data, they shouldn't be a convergence problem, the other customers should be able to smooth them out. I suspect your problem is elsewhere.

I suggest you share a sample of your data that has the same issue here, so others can help trying to solve it.