Biased or Debiased: Polarization-aware Embedding Learning from Social Media Knowledge Graph

Nowadays social media has become a major information source for understanding various aspects of social activities, such as political polarization and shopping preferences. To use social media data with machine learning algorithms, the data can be transformed into graphs, from which graph embeddings can be learned. However, it has been shown that embeddings usually inherit the social bias prevalent in the data. Sometimes such bias is necessary, and sometimes they are undesirable. Previous work that deals with embedding bias usually tries to neutralize embeddings with regard to a bias measure. In this paper, we propose an embedding bias model and a control method, which can reduce or enlarge the bias by changing the value of one parameter. In addition to explicit bias scores, we also test our method in two downstream tasks, namely, party support prediction and tweet recommendation. Through experimental evaluations, we show that when we reduce the embedding bias, the task irrelevant to the sensitive attribute can be improved. On the other hand, when we enlarge the bias, the task relevant to the sensitive attribute can be improved.