SLIDE 1
how algorithmic confounding in recommendation systems increases homogeneity and decreases utility
Allison J.B. Chaney Princeton University
In collaboration with Brandon M. Stewart and Barbara E. Engelhardt
how algorithmic confounding in recommendation systems increases - - PowerPoint PPT Presentation
how algorithmic confounding in recommendation systems increases - - PowerPoint PPT Presentation
how algorithmic confounding in recommendation systems increases homogeneity and decreases utility Allison J.B. Chaney Princeton University In collaboration with Brandon M. Stewart and Barbara E. Engelhardt Simulation Setup Simulation Setup
SLIDE 2
SLIDE 3
Simulation Setup
SLIDE 4
Simulation Setup
alternative realities content filtering social filtering matrix factorization popularity random ideal “world”
SLIDE 5
Jaccard Index
|A ) = ∩ B| ) =
SLIDE 6
Jaccard Index
) = J(A, B) = |A ∩ B| |A ∪ B|
|A ) = ∩ B| ) =
SLIDE 7
100
iteration
SLIDE 8
100
iteration
SLIDE 9
100
iteration
SLIDE 10
Claim 1: The recommendation feedback loop causes homogenization of user behavior.
SLIDE 11
content MF popularity random social −0.6 −0.4 −0.2 0.0 −0.6 −0.4 −0.2 0.0 −0.6 −0.4 −0.2 0.0 −0.6 −0.4 −0.2 0.0 −0.6 −0.4 −0.2 0.0 −0.25 0.00 0.25 0.50
utility relative to ideal change in Jaccard index
SLIDE 12
content MF popularity random social −0.6 −0.4 −0.2 0.0 −0.6 −0.4 −0.2 0.0 −0.6 −0.4 −0.2 0.0 −0.6 −0.4 −0.2 0.0 −0.6 −0.4 −0.2 0.0 −0.25 0.00 0.25 0.50
utility relative to ideal change in Jaccard index
SLIDE 13
content MF popularity random social −0.6 −0.4 −0.2 0.0 −0.6 −0.4 −0.2 0.0 −0.6 −0.4 −0.2 0.0 −0.6 −0.4 −0.2 0.0 −0.6 −0.4 −0.2 0.0 −0.25 0.00 0.25 0.50
utility relative to ideal change in Jaccard index
SLIDE 14
content MF popularity random social −0.6 −0.4 −0.2 0.0 −0.6 −0.4 −0.2 0.0 −0.6 −0.4 −0.2 0.0 −0.6 −0.4 −0.2 0.0 −0.6 −0.4 −0.2 0.0 −0.25 0.00 0.25 0.50
utility relative to ideal change in Jaccard index
SLIDE 15
Claim 2: Users experience losses in utility due to homogenization effects; these losses are distributed unequally.
SLIDE 16
Gini Coefficient
) =
G(A, B) = A A + B
line of equality items ordered by popularity popularity of items
A B
item popularity curve (usually long tail)
SLIDE 17
Gini Coefficient
) =
G(A, B) = A A + B
line of equality items ordered by popularity popularity of items
A B
item popularity curve (usually long tail) maximal equality
G ∈ [0, 1]
maximal inequality
SLIDE 18
SLIDE 19
Claim 3: The feedback loop amplifies the impact of recommendation systems
- n the distribution of item consumption.
SLIDE 20
Why do we need to think about algorithmic confounding?
SLIDE 21
Why do we need to think about algorithmic confounding?
better evaluation of recommendation systems
SLIDE 22
Why do we need to think about algorithmic confounding?
better evaluation of recommendation systems understand the impacts on human behavior
SLIDE 23