Predicting the Popularity of GitHub Repositories | Hacker News Reader

Predicting the Popularity of GitHub Repositories

arxiv.org

36 pointsnextjj6 comments

Umm they use the first 6 months of data in the model. How is it surprising that trajectory of stars in first 6 months predicts the rest of the trajectory.....

> when newcomers are not considered, there is a very strong correlation between predicted and real rankings

I'm having trouble parsing the meaning of this. It seems like their data can predict the popularity but only on existing repositories in which they can look at the first 6 months, calculate the popularity and see if it matches up. Am I understanding that correctly? Seems...weird. Can this be used in any way for new projects?

For instance my msngr.js[1] library initially had like 3 stars from friends then it was posted to HN and within 2 weeks had about 150 then it slowly climbed to what it is today over a year or so and lately it's kinda stalled. I'm not sure their data would predict anything correctly about my repository and I would imagine a large majority are more like mine where you might have a handful of users and some pop when being on the front page of something like HN but then traffic dies back down.

[1] https://github.com/KrisSiegel/msngr.js

So... They found that the number of stars is autoregressive? What's the punchline?

IIUC the utility is that github or someone could use the arrival of stars to predict which repos are likely to attract more future stars. This could be used for ranking up-and-coming new repos worth attention.

From what I remember of reading the paper yesterday, they simply estimated the coefficient of various lag terms. That might help if you wanted to estimate how many stars a project will have next week (because of some prediction market?). For the task of predicting which projects will attract more stars, the fitted model is largely unnecessary -- more means more.

Have data; find paper.

Reply on news.ycombinator.com