Auto insight allows you to easily identify how one state of a variable differs from the others. For example if a model contains a variable Purchase with states True and False, we might be interested in how Purchase=True differs from Purchase=False. For a mixture model, one cluster might perform better than others, so we might be interested in how this cluster differs from the others.
The auto insight viewer allows us to easily identify the key differentiators, which can be sorted by Difference in order to detect the largest patterns, or lift to identify anomalies.
Once the first key differentiating states have been identified, you can then click on one of these states in a pane in the viewer, and the calculations will be repeated, this time with additional evidence set on that state.
Drilling down is a little bit like a decision tree, however it is dynamic and the calculations are graph based.
By drilling down you can approximate/convert a probabilistic model into a partition of data much like a WHERE clause in SQL, which can be useful when trying to provide multivariate insight. The number of levels you drill down, depends on what you are trying to optimize. You might stop when the target probability no longer increases, or when the selection likelihood is too small.
When Use network evidence is selected, any evidence set on the network before opening the auto insight viewer will be used in the subsequent calculations.
When a state is clicked in a pane, the likelihood is displayed which includes the path clicked to get to this state, and any network evidence.