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 alternative realities “world” content filtering social filtering matrix factorization popularity random ideal

Jaccard Index ) = ) = | A ∩ B |

Jaccard Index ) = ) = | A ∩ B | J ( A, B ) = | A ∩ B | ) = | A ∪ B |

100 0 iteration

100 0 iteration

100 0 iteration

Claim 1: The recommendation feedback loop causes homogenization of user behavior .

change in Jaccard index content MF popularity random social 0.50 0.25 0.00 − 0.25 − 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 utility relative to ideal

change in Jaccard index content MF popularity random social 0.50 0.25 0.00 − 0.25 − 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 utility relative to ideal

change in Jaccard index content MF popularity random social 0.50 0.25 0.00 − 0.25 − 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 utility relative to ideal

change in Jaccard index content MF popularity random social 0.50 0.25 0.00 − 0.25 − 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 utility relative to ideal

Claim 2: Users experience losses in utility due to homogenization effects; these losses are distributed unequally .

Gini Coefficient A G ( A, B ) = line of equality A + B popularity of items item popularity curve (usually long tail) A ) = B items ordered by popularity

Gini Coefficient A G ( A, B ) = line of equality A + B popularity of items item popularity curve (usually long tail) A G ∈ [0 , 1] ) = B maximal maximal equality items ordered by popularity inequality

Claim 3: The feedback loop amplifies the impact of recommendation systems on the distribution of item consumption .

Why do we need to think about algorithmic confounding ?

Why do we need to think about algorithmic confounding ? better evaluation of recommendation systems

Why do we need to think about algorithmic confounding ? better evaluation of recommendation systems understand the impacts on human behavior

Why do we need to think about algorithmic confounding ? better evaluation of recommendation systems understand the impacts on human behavior design better systems to 
 increase fairness and social welfare