SEKA: Seeking Knowledge Graph Anomalies
Asara Senaratne · 2023
Knowledge Graphs (KGs) form the backbone of many knowledge dependent applications such as search engines and digital personal assistants. KGs are generally constructed either manually or automatically using a variety of extraction techniques applied over multiple data sources. Due to the diverse quality of these data sources, there are likely anomalies introduced into any KG. Hence, it is unrealistic to expect a perfect archive of knowledge. Given how large KGs can be, manual validation is impractical, necessitating an automated approach for anomaly detection in KGs. To improve KG quality, and to identify interesting and abnormal triples (edges) and entities (nodes) that are worth investigating, we introduce SEKA, a novel unsupervised approach to detect anomalous triples and entities in a KG using both the structural characteristics and the content of edges and nodes of the graph. While an anomaly can be an interesting or unusual discovery, such as a fraudulent transaction requiring human intervention, anomaly detection can also identify potential errors. We propose a novel approach named Corroborative Path Algorithm to generate a matrix of semantic features, which we then use to train a one-class Support Vector Machine to identify abnormal triples and entities with no dependency on external sources. We evaluate our approach on four real-world KGs demonstrating the ability of SEKA to detect anomalies, and to outperform comparative baselines.