Investigate regularization. Currently, density ratio parameters are regularized to zero, which implies large denominator mass compared to numerator mass, which seems, intuitively, a much stronger assumption than regularizing to one, which yields no preference for either dataset.
See this link for a potential solution: https://gist.github.com/thomvolker/afdca959f92f29b410ca05c861caae75.
The question is whether this works accordingly when performing cross-validation. Also, check whether this approach accommodates the fact that the starting point is one, rather than zero, when estimating the density ratio parameters.
Investigate regularization. Currently, density ratio parameters are regularized to zero, which implies large denominator mass compared to numerator mass, which seems, intuitively, a much stronger assumption than regularizing to one, which yields no preference for either dataset.
See this link for a potential solution: https://gist.github.com/thomvolker/afdca959f92f29b410ca05c861caae75.
The question is whether this works accordingly when performing cross-validation. Also, check whether this approach accommodates the fact that the starting point is one, rather than zero, when estimating the density ratio parameters.