Data Science of the KDD ‘14 Review Process Jure Leskovec (Stanford) and Wei Wang (UCLA) Joint work with Jason Hirshman and David Zeng (Stanford)
Data Science of the KDD ‘14 Review Process
Jure Leskovec (Stanford) andWei Wang (UCLA)
Joint work with Jason Hirshman and David Zeng (Stanford)
KDD 2014 Research Track Statistics
KDD 2014 Program
Largest KDD program ever:• 151 research papers (20% growth over KDD’13)• 43 industry & govt. papers (30% growth)• 26 workshops (75% growth)• 11 tutorials (83% growth)
Program highlights:• Paper spotlights early morning (8:15am)• Oral presentations (Mon-Wed)• Posters at the reception (Tue night)
KDD 2014 Research Track
• 1036 submissions from 2600 authors– 42% increase over KDD ’13
• 151 papers:– Acceptance rate
14.6%
20002002
20042006
20082010
20122014
20160
200
400
600
800
1000
1200
KDD year
Num
ber o
f sub
miss
ions
KDD Reviewing Process
46 Senior PC members + 340 PC members• 2971 reviews in total
(Rough) Acceptance rule: • Raw review score AND Standardized review score AND Raw
meta-review AND Standardized meta-review score ≥ Weak Accept
• 110 papers matched (immediate accepts)• Remaining papers were discussed with meta-reviewers and
final decisions were made
Submissions per Country
Acceptance Rate per Country
Acceptance by Subject Area
Predicting Paper AcceptanceFeatures Used AccuracyRandom Guessing 0.50Paper Abstract 0.57
Author Status (Past paper counts) 0.64
Author Status (DBLP graph connectivity) 0.61
Author Status (Counts + Graph) 0.65
Reviewer (Similarity, Graph distance to authors) 0.60
All (Abstract, Author Status, and Reviewer) 0.65
Predicting Paper Acceptance from the Review Text
Features Used Paper: Accepted?
Review: Score > 0?
Random Guessing 0.50 0.50
Review Text 0.68 0.72
Review Text + Numeric Score (Novelty, Presentation) 0.77 0.77
Human Reading of Review Text 0.88 0.73
I’m submitting a paper:What correlates with acceptance?
Academia + Industry Papers do Better
Submissions per Author: 5 is best!
No benefit in submitting >5 papers!
Having more authors (seems to) help
It is the most experienced author that matters!
What insights can we gain on the review process?
Most reviews are Weak Rejects
More granularity is needed at the Weak Reject / Weak Accept level
Revi
ew a
gree
s with
the
final
out
com
e
Review length is a good determinant of a review’s influence/quality
Revi
ew a
gree
s with
the
final
out
com
e
Shorter reviews are used for clear accepts and rejects
Never review co-author’s papers
The Curse of the Review Submission Deadline
Over 50% reviews submitted in the last 5 daysOver 20% reviews submitted in the last 24 hours
10% of reviewssubmitted late
Ratings increase near the deadline
Weak Rejects increase while
Rejects decrease
Reviews submitted late are less likely to agree with final outcome
Late reviews are shorter
Review quality drops: Accuracy of predicting score from review text
Conclusions• To get your papers accepted to KDD:– Collaborate in multidisciplinary teams– Have a senior author on board– Do not submit more than 5 papers
• To improve KDD community standards:– Avoid Weak Reject/Weak Accept scores– Write longer and clearer reviews– Submit reviews early!