Counterfactual Explanations for Recommendation Bias

Panayiotis Tsaparas

University of Ioannina

Athena Research Center - Archimedes Unit

Greece

Evaggelia Pitoura

University of Ioannina

Athena Research Center - Archimedes Unit

Greece

Leonidas Zafireiou

University of Ioannina

Greece

Recommendation Systems

• Recommendation systems are successful in assisting us in a variety of tasks

​

• We want to provide explanations for biases in recommendations

Job offers based on gender

• However, recommenders may incorporate biases in the training data and lead to unfair recommendations

Explanations

• Counterfactual explanations: The smallest change in the input data to achieve a desired output

​

• Typically, instance-level explanations

​

• To explain bias, we need group-level explanations

Background

Bias Definitions

• Individual User Bias

​

​

• User Group Bias

​

​

• Individual Item Bias

​

​

• Item Group Bias

​

​

Problem definition

Recommendation algorithm

Individual Biases

Individual User Bias

Individual Item Bias

Group biases

Group biases

User group bias

Item group bias

Thank you!

Questions?