Table 2.

Three versions of the ISMB algorithm.

Acronym	Algorithm	Phase 1	Phase 2	Prediction
ISMB	Instance-specific Markov blanket	Is non-instance-specific Uses greedy hill-climbing search Uses phase 1 score	Is instance-specific Uses best-first search Uses phase 2 score	By model averaging over models selected in phase 1 and phase 2
ISMB-MS	Instance-specific Markov blanket - Model Selection	Same as ISMB	Same as ISMB	Based on the highest scoring model from models found by ISMB
NISMB	Non-instance-specific Markov blanket	Same as ISMB	Is non-instance- specific Uses best-first search Uses phase 1 score	By model averaging; number of selected models is the same as in ISMB