Knowledge Graph Machine Learning

Knowledge graphs have a number of features that make them desirable in the context of Machine Learning algorithms. They are self-descriptive, efficient, and human-interpretable. They may be used as inputs to Machine Learning algorithms, or produced as outputs of algorithms.

Knowledge Graphs can be used as a highly compact and flexible format for storing data and feeding it into a Machine Learning algorithm. This data structure is self-descriptive, meaning that all information needed to interpret the data is captured in the graph; it is schema-less. This makes it easy to combine multiple unstructured or semi-structured data sources as inputs. Knowledge graphs are also computationally efficient for storing and querying purposes. Drug discovery is a problem well-suited for this methodology, since it requires integrating as much data as possible from various research labs and studies to try and understand the relationships between genes, drugs, and diseases.

Knowledge Graphs also allow for better explainability of ML results by mapping predictions to relevant nodes in the graph, which is much more human-readable than a trained neural network. Knowledge graphs can be produced as outputs from feature extraction algorithms in text and images. These algorithms might identify a number of objects (nodes) in an image or subjects in text, and then determine the relationships (edges) between those objects, both active and static. A Natural Language Processing algorithm would do the same, using text as an input.