Order Scoring Test Result Observations
Posted: Mon Aug 25, 2014 3:40 pm
- Test #1's log score for the dataset
X0 X1 X2
0 0 0
comes out to a positive value (4.44089209850063e-16).
I calculated this by hand (using BDeu metric's priors) and also got a positive log score (essentially slightly greater than 0). I am not too sure if this should be a concern. - When generating a dataset, it is often difficult to reflect a particular total order.
For example, the scores for two binary nodes is typically identical between X1<X2 and X2<X1.
Also, if I imagine an length-n chain X1->X2->...->Xn, where all conditional probabilities are uniform (e.g. P(Xi=0|X{i-1}=0)=P(Xi=1|X{i-1}=0)=...=0.5 ), it looks like knowing X{i-1} doesn't seem to affect Xi. - Anyway, I think the scores are correct (based on some rough calculations).