4

I got my Bachelor's degree in Statistics and I'm starting graduate school to study Computer Science (CS) with an emphasis on Data Mining. One of the big differences that I've noticed between the Statistics and the ML communities is that Statistics tends to publish in journals, while CS tend to publish at conferences.

I've heard people say that conference presentations are preferred to publishing in journals because "things move so fast" in Machine Learning research. However, if you actually look reputable statistics journals are published very frequently.

The American Statistical Association publishes Statistical Analysis and Data Mining every other month and the Journal of Computational and Graphical Statistics is published quarterly. It looks like they both also have pretty quick turnaround times. The International Statistical Institute even has a journal that was started specifically as a "rapid communication research journal" that published papers monthly.

So, if presenting at conferences isn't actually faster than publishing in a credible statistics journal, how did this division between the two communities start?

7
  • ML communities do have a lot of journals, and on the other side, even the stats communities have a lot of conferences
    – Dawny33
    Commented Dec 4, 2015 at 13:08
  • @Dawny33 I know that there are statistics conferences, like the JSM, and I know that there are ML journals, like the Journal of Machine Learning Research. Correct me if I'm wrong, but it seems like the preferred method of doing things in the ML community is to present at conferences. It seems like there is more weight attached to a conference presentation as opposed to a publication. Am I wrong?
    – j.jerrod.taylor
    Commented Dec 4, 2015 at 13:25
  • Yeah, I agree. The fact that conferences(ML) > conferences(stats) make you think that that the interest in journals is less in the ML community. The huge number of conferences is due to the wide application of ML in the industries and companies. So, they are mostly abt the applications, etc.
    – Dawny33
    Commented Dec 4, 2015 at 15:21
  • @Dawny33 Agree... inclination is that the answer resides more in who is using and developing in each arena. Stats being primarily academic and ML being primarily industry. Also, the peer review that accompanies a journal may be unnecessary (already proven in the market place) or unwanted (IP).
    – Dave
    Commented Dec 4, 2015 at 17:56
  • Off topic. This should goto academia.overflow.
    – SmallChess
    Commented Dec 5, 2015 at 4:48

1 Answer 1

1

This came up just the other day in a class - journal vs conference publications.

Background; pros and cons:

I came to CS from neuroscience, a field (like most?) in which journals were the primary currency and some of the larger conferences accepted posters / talks but with only a minor vetting process. Conferences were often used to present preliminary work and get feedback on methodology, journal articles were for the meat of the work. I had many conference posters (and one or two journal papers) on my CV when i applied to CS programs and was surprised at how professors were impressed by that - it seemed pretty common for people in neuro to have a good number of conference posters/talks.

Conference papers have low latency of acceptance/presentation. Typically deadlines are 4-7 months out from the conference, which tends to be quite a bit faster than journal article acceptances / publications. In my relatively short career, i've worked on papers (in neuroscience) that took 2+ years to publish, from intial submission to revision rounds to acceptance and publication. Conferences also tend to give your work more visibility - I recently presented my recent research at an SC15 workshop and had more feedback and potential collaborations come out of it than i ever did from my 1st-author journal article or from poster sessions in neuro conferences.

Journal articles allow for considerably greater depth and treatment of the topic at hand and are much more useful for reproducibility and understanding. Conference papers tend to have hard page limits and if the work is substantial enough, you can be forced to gloss over detail and have the potential to not ever publish follow up papers to explain the missing details.

A good issue around this, and this is what we discussed in class, was Google's Spanner paper, presented at OSDI. The paper is groundbreaking and incredibly impressive work, but due to the limitations of the format (short paper, again), much of the detail and nuance needed to understand the work is left out. The paper even mentions that a follow up paper would be forthcoming - three years later and it hasn't materialized. For such a huge piece of work that encompassed years of principle engineer time at an org like Google, it was probably better suited to a 30~ page journal article, but it may not have had the same impact as it did being presented at OSDI - one of the premier systems conferences.

your question

To answer the question at hand, (why is it like this), many people have written about it (at conferences and journals!). One longer discussion can be found here.

Perhaps it is as simple as "this is how it has been for a long time and change is hard" - not particularly satisfying, but might explain a great deal of the variance here.