peer review in the liquidpub project
DESCRIPTION
Slides presented at the 2nd Snow Workshop (http://wiki.liquidpub.org/mediawiki/index.php/Second_Workshop_on_Scientific_Knowledge_Creation%2C_Dissemination%2C_and_Evaluation)TRANSCRIPT
Reviewing Peer Review
Metric DimensionsQuality
FairnessEfficiency
Statistics
Kendall Distance
Divergence
Disagreement
Biases
Robustness
Unbiasing Effort invariant alternatives
Effort vs. quality
Min. criteria
Quality-related Metrics: real vs. ideal
• Real peer review ranking vs. ideal ranking▫ Ideal ?
Subjective vs. Objective But each process could/should define approximate
indicators of quality like: citations, downloads, community voting, success in a second phase, publication, citations, patents…
• IF an approximate ideal ranking is available we can measure the difference in various ways, e.g▫Kendall distance / Kendall rank correlation▫Divergence metric
Divergence Metric
©√
Nt = 1/nNormalized t
Nt = n/n
N-D
iver
gen
ceNdiv(1,n, C) = n-1/n
©√
€
NDivρ iρ a (t,n,C ) = pt (i)wii=0
t
∑e.g.
NDivρ iρ a (1,n,C ) = p1(0)1
1 ⎛ ⎝
⎞ ⎠+ p1(1)
0
1 ⎛ ⎝
⎞ ⎠
=n −1
n
1
1 ⎛ ⎝
⎞ ⎠
....
pt (i) =CitCt−i
n−t
Ctn ;wi =
t − i
t
When the second ranking is random, we have:
indipendent
correlated
inv. correlated
prior vs. after discussion
€
NDiv(53,206,206) = 0, 36 ca. 74 (36%) contributions have been effected by the discussion phase
Results: peer review ranking vs. citation count
6
Div
Normalized t
Fairness
•Definition: A review process is fair if and only of the acceptance of a contribution does not depend on the particular set of PC members that reviews it
•The key is in the assignment of a paper to reviewers: a paper assignment is unfair if the specific assignment influences (makes more predictable) the fate of the paper.
Computed Normalized Rating Biases
C1 C2 C3 C4
top accepting 2,66 3,44 1,52 1,17
top rejecting -1,74 -2,78 -2,06 -1,17
> + |min bias| 13% 5% 9% 7%
< - |min bias| 12% 4% 8% 7%
C1 C2 C3 C4
Unbiasing effect (divergence) 13% 9% 11% 14%
Unbiasing effect (reviewers affected) 10 16 5 4
Disagreement metric
•Through this metric we compute the similarity between the marks given by the reviewers on the same contribution.
•The rationale behind this metric is that in a review process we expect some kind of agreement between reviewers.
10/13
Disagreement vs number of reviews
The road aheadReal-Time accuracy estimation
Speed Ranking
Ranking vs. marking
Thank you