George Berry  

@george_berry

Sociologist working at Civis. Interests: experiments, social influence, social networks, train memes. he/him.

Brooklyn, NY
Vrijeme pridruživanja: lipanj 2011.

Tweetovi

Blokirali ste korisnika/cu @george_berry

Jeste li sigurni da želite vidjeti te tweetove? Time nećete deblokirati korisnika/cu @george_berry

  1. Prikvačeni tweet
    31. sij

    Are you studying social relationships online? Are you using predictions for node attributes? 👇You may want to check out our new preprint👇

    Prikaži ovu nit
    Poništi
  2. proslijedio/la je Tweet
    prije 5 sati

    In the past, I've assigned this very useful paper on ad measurement: But I think it is getting a bit out-of-date, including some of its (well-founded) pessimism. What's the best updated overview of learning about ad effects?

    Poništi
  3. Poništi
  4. proslijedio/la je Tweet
    3. velj

    I love Super Bowl statistics. (“No team that has failed to convert at least 35% of their 3rd downs has gone on to win” etc.) They are the best introduction to p-hacking that can be found outside major journals.

    Poništi
  5. proslijedio/la je Tweet
    2. velj

    One of the nice things about Andrew Gelman's blog is how benign posts routinely produce deep debates about foundations of statistics. This time, the concept of "calibration" (see comments): My attitude (new box in 2nd ed of my book):

    Prikaži ovu nit
    Poništi
  6. 1. velj

    Catching up on the SB50 fallout and I wonder if people tend to perceive density and affordability as tradeoffs. It seems like density and livability are seen as tradeoffs.

    Poništi
  7. 1. velj

    No, autocorrect, I did not mean to type Iowa cactuses

    Poništi
  8. proslijedio/la je Tweet
    1. velj

    Nigerian immigrants are by far the most educated immigrant group in the US. Over 20% have *graduate* degrees. But... They're also black. So due to the "travel ban" Nigerians will no longer be allowed to get green cards. The ban was never about security.

    Prikaži ovu nit
    Poništi
  9. 31. sij
    Poništi
  10. 31. sij

    Relatedly I’d love to see a full GNN approach to unbiased estimation of network quantities. If you want to team up on this lmk!

    Poništi
  11. 31. sij

    Since this is a brand-new preprint, we'd love to hear any comments you have, please reach out! Big shoutout to my co-authors , , , and .

    Prikaži ovu nit
    Poništi
  12. 31. sij

    Also worth mentioning: similar reasoning can be applied to any task which requires aggregating predictions in a weighted way. For instance, if you're estimating how much hate speech people see, and a model tends to make larger errors on the posts people see the most.

    Prikaži ovu nit
    Poništi
  13. 31. sij

    The main suggestions we have for researchers are: - Include network information in classifiers (this is a must!!) - Try out the ego-alter modeling strategy we propose in the paper, which in simulations studied here performs well.

    Prikaži ovu nit
    Poništi
  14. 31. sij

    When thinking about this problem, we hypothesized that classification errors would always reduce homophily. But we found that when model errors are correlated with node degree, the opposite can be true! Classification can create an *upward bias* in homophily.

    Prikaži ovu nit
    Poništi
  15. 31. sij

    Additionally, node-level model performance measures (AUC, accuracy) aren't very informative about homophily bias. Small differences in e.g. AUC relate to large differences in bias. The key is that the relationship between errors and network are important.

    Prikaži ovu nit
    Poništi
  16. 31. sij

    From this, it follows that it's *crucial* to include network information in the classifier. Even if you have a random sample of nodes/edges, a classifier without network information can introduce large biases. Check out the "node (no network)" model here. 👇

    Prikaži ovu nit
    Poništi
  17. 31. sij

    Common homophily measures are a dyadic prediction problem. This means if you use node-level predictions, errors are multiplied along dyads. This can introduce lots of bias! We show you can express bias as a sum of residuals over the dyads.

    Prikaži ovu nit
    Poništi
  18. proslijedio/la je Tweet
    29. sij

    In the middle of the last century, main streets became highways. There were many costs, which politicians of that era ignored. Major cities around the world, and now two in the US (NYC and as of today SF) are reversing that decision and reaping the economic and social benefits.

    Prikaži ovu nit
    Poništi
  19. proslijedio/la je Tweet
    29. sij

    Starting today, San Francisco's main civic boulevard puts people first. Market Street is officially car-free!

    Poništi
  20. 29. sij

    I am convinced housing discourse breaks people's brains faster than any other issue.

    Poništi

Čini se da učitavanje traje već neko vrijeme.

Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.

    Možda bi vam se svidjelo i ovo:

    ·