Truthful Illustration Studying for Heterogeneous Info Networks
Authors: Ziqian Zeng, Rashidul Islam, Kamrun Naher Keya, James Foulds, Yangqiu Song, Shimei Pan
Summary: Lately, a lot consideration has been paid to the societal impression of AI, particularly issues concerning its equity. A rising physique of analysis has recognized unfair AI techniques and proposed strategies to debias them, but many challenges stay. Illustration studying for Heterogeneous Info Networks (HINs), a basic constructing block utilized in advanced community mining, has socially consequential purposes resembling automated profession counseling, however there have been few makes an attempt to make sure that it is not going to encode or amplify dangerous biases, e.g. sexism within the job market. To deal with this hole, on this paper we suggest a complete set of de-biasing strategies for truthful HINs illustration studying, together with sampling-based, projection-based, and graph neural networks (GNNs)-based methods. We systematically research the conduct of those algorithms, particularly their functionality in balancing the trade-off between equity and prediction accuracy. We consider the efficiency of the proposed strategies in an automatic profession counseling utility the place we mitigate gender bias in profession suggestion. Primarily based on the analysis outcomes on two datasets, we establish the best truthful HINs illustration studying methods beneath totally different situations