What’s the main problem with existing AI models and development?
Interpretability. This stops AI from being widely adopted in industries where it’s important to be able to explain the model inferences. This has led to many R&D on explainable AI!
What’s this paper about?
This paper is about Fujitsu Laboratories achieving the world’s first “trustworthy and explainable AI” through using knowledge graphs to represent and aggregate knowledge and data. We can use knowledge graphs to explain and and perform reasoning.
What is explainable AI?
AI that can explain the reasoning behind its prediction. Current R&D has been focusing on determining which input data is the deciding factor in an AI inference.
How did Fujitsu Lab configure the explainable AI using knowledge graphs?
The overall configuration is shown in the figure below. The training data is graph-structured data in deep tensor and the outputs are inference result AND inference factors (reasons). The knowledge graph gives the basis for the inference result.
What is a knowledge graph?
It’s used to represent knowledge in the form of connections between concepts. Entities are nodes and the relationship between entities are edges.
How to build a knowledge graph?
As shown in the figure below, there are two components:
Consume data to derive the entities and relationships
Come up with an ontology to associate between these concepts and entities. The ontology allows you to perform reasonings as you get to determine the connections between two entities. The ontology is usually created by domain experts.
How can KG be apply to finance?
At current state, banks can provide reasons for predicted output only at the input feature level, i.e. identifying which features were relied on heavily by the model to make the prediction. This means that we are not at the stage where we can understand the model’s thought process of going from input to output.
One application at Fujitsu Lab is using KG to determine growing companies by analysing their credit scores. A financial knowledge graph is a knowledge base in graphical form that’s built using publicly available corporate numbers and open data.
Another application is evaluating loan applications and credit risk.