A closed dot connected to a set of open dots indicates that the condition aligned with the closed dot significantly "dominates" all of the conditions aligned with the open dots. The meaning of "domination" is context-dependent (e.g., significantly better accuracy, significantly lower error rate, significantly higher count and so on). This visualization convention was first introduced here.
A more elaborate usage case:
See demo.py for how to use this code.
Requirements: matplotlib and pandas.
The demo uses also seaborn (for barplots and boxplots), statsmodels, and numpy. The demo data (Eysenck, 1974) is from David Howell's website.