Hybrid Link Prediction for Competitor Relationships

Competitor relationships are integral to many important financial applications. Example use cases include understanding regulatory impacts, investing in new business areas, and building economic models. Competitor relationships can be defined based on several aspects, including valuations and asset returns, industrial processes, or offerings of products and services. Determining these relationships is often challenging due to the diverse and complex interactions between companies which must be mined from vast datasets with varying degrees of credibility. In this paper, we approach this problem by constructing a hybrid knowledge graph capturing financial relationships and applying a link prediction model to identify missing competitor relationships. Knowledge graphs are a popular knowledge representation choice for capturing entities and the relationships between them. Knowledge graph construction typically uses only a single type of input data, such as relationships mined from text using information extraction techniques or curated relationships from relational databases. In contrast, for the FEIII Challenge, we are provided with several sources of relationships from different types of input, including expert judgments, mined relationships, and statistical features. Our approach creates a hybrid knowledge graph that includes relationships derived from three very different types of data in a single knowledge graph. We construct a hybrid knowledge graph using data provided for the FEIII Challenge and one additional source, the webpages of companies included in the challenge. The first data source we use are expert judgments curated by the Thomson Reuters Data Fusion (TRDF) platform. The second data source we are provided in the challenge are relationships extracted from text found in SEC filings. Finally, we introduce a third set of statistical signals, derived primarily from collecting webpages of the companies in the