Reducing Bias In Link Prediction

Link prediction, the task of infering completing knowledge graphs by predicting missing relationships between entities has been shown to enhance performance for further downstream tasks like decision-making. However, link prediction methods are susceptible to inheriting biases present in small and incomplete datasets are which serve as training data, potentially resulting in skewed predictions. This working paper investigates approaches for mitigating this bias on a specific real-world example in a medium-sized enterprise with only limited data capabilities.