MultiGraph Attention Network for analyzing Company Relations

When analyzing companies in financial markets, it is essential to identify those companies that share similar characteristics in order to assess their relative strengths and weaknesses. This challenging task requires representing the rich set of information associated with companies and the complex interrelations between them in a form that is amenable to pattern recognition. We present here a new deep representation learning method that encodes the network graph of companies in a low-dimensional embedding space, preserving its topological structure. Our solution employs a number of neural attention mechanisms that adaptively aggregate information over company node neighborhoods in a multi-dimensional edge setting. The learned company embeddings are transferable and can be fine-tuned for a wide range of analytical tasks. We demonstrate improvement over state-of-the-art solutions and illustrate the efficacy of our method for financial analysis tasks such as industry classification, peer group identification, credit rating anomaly detection and visualization of company relations.