First-timer at the INSNA Sunbelt Conference, this time in Edinburgh. It wasn’t ideal to present in the very last session on the last day of the conference, so thank you to those who turned up, and the numbers were indeed better than expected.
Some major differences of this presentation with the extremes reading group:
- This is (obviously) more network focused, with the main idea being that most networks are at best partially scale-free, and the use of the mixture distribution facilitates determining such partial scale-freeness.
- The distribution has been fitted to the CRAN dependency network at different time points, to show some sort of stability of the results over time. The tail is steadily heavy ($\xi>0$) but lighter than that implied by the power law ($1/(\alpha-1)$).
- Point 2 raises the question: what processes lead to this kind of stability? I would postulate a preferential attachment model with fatigue or attenuation, meaning that the attachment probability doesn’t increase linearly once the degree reaches certain level.
Also some useful comments from the audience:
- Looking at
importsonly might not be the best for CRAN dependencies; perhaps tryimports+depends, ordependsalone, which has a stronger meaning thanimports. - A latent attribute / variable for each node to represent the usefulness and/or the general applicability of the package would be useful.
Hopefully I will have something more concrete over the next few months, as I do want to go to Paris (where the conference takes place next year)!