In a significant stride towards enhancing knowledge graph (KG) technology, a team of researchers led by CHEN Heng from Dalian University of Foreign Languages and WANG Siyi from Dalian Maritime University has developed a novel model that promises to improve the accuracy of predicting missing information in knowledge graphs. This research, published in the journal ‘Jisuanji gongcheng’ (translated to ‘Computer Engineering’), integrates quaternion into capsule neural networks to better complete missing fact triplets in knowledge graphs.
Knowledge graphs are essentially structured representations of information, using a format that describes relationships between entities in the real world, such as (subject, predicate, object). For instance, in a maritime context, a knowledge graph might represent relationships between ships, ports, and cargo types. The challenge lies in completing these graphs when some information is missing.
The researchers’ innovative approach involves using hypercomplex embedding to encode the structure of these triplets, replacing traditional real-valued embedding. This method captures global features of the triplets more effectively. As CHEN Heng explains, “By using quaternion embedding of entities and relationships as the input of the capsule network, we can effectively complete the missing triplets in the knowledge map and improve the prediction accuracy.”
The implications for the maritime sector are substantial. Knowledge graphs are already used in various applications, from logistics and supply chain management to route optimization and predictive maintenance. Enhancing the accuracy of these graphs can lead to more efficient operations, better decision-making, and ultimately, cost savings.
For example, a more accurate knowledge graph could help shipping companies predict maintenance needs more precisely, reducing downtime and repair costs. It could also improve route planning by better predicting weather patterns, port congestion, and other variables that affect shipping routes.
The researchers tested their model on two datasets, WN18RR and FB15K-237, and found significant improvements over the existing CapsE model. On the WN18RR dataset, their model achieved a 3.2 percentage point higher accuracy in Hit@10 and a 5.5% higher reciprocal average ranking. On the FB15K-237 dataset, the improvements were 2.5 percentage points in Hit@10 and 4.4% in reciprocal average ranking.
WANG Siyi highlights the practical benefits: “Our model can effectively predict the missing fact triplets in a knowledge graph, which can be a game-changer for industries relying on accurate and comprehensive data.”
As the maritime industry increasingly adopts digital technologies, the demand for accurate and efficient knowledge graphs is set to grow. This research not only advances the field of knowledge graph completion but also opens up new opportunities for maritime professionals to leverage these advancements for better operational efficiency and strategic decision-making.
In the ever-evolving landscape of maritime technology, this breakthrough represents a significant step forward, offering a glimpse into the potential of advanced data structures and machine learning techniques to transform the industry.