Translation Embeddings for Knowledge Graph Completion in Consumer Banking Sector

Knowledge graphs have shown many successful applications such as recommendation. One essential challenge of knowledge graph is link prediction, which is to estimate the probability of a link between two nodes based on the existing graph. Most state-of-art models to solve this problem are built using embeddings. Experiments of previous works have been conducted on WN18 and FB15K datasets to compare the performance. However, both datasets have significantly different properties compared with financial data on which there is no benchmarking of link prediction models. In this paper, we run extensive experiments of recent models on real financial data, compare their performance deeply, and show the usage of a completed knowledge graph in consumer banking sector.