Data is the fuel for models but you may have witnessed situations where there is no data but solely a domain expert that can very well describe or even predict “the situation” given the circumstances. I will summarize the concepts of knowledge-driven models in terms of Bayesian probabilistic, followed by a hands-on tutorial to demonstrate the steps of converting an expert’s knowledge into a Bayesian model with the goal to make inferences. I will use the Sprinkler system to conceptually explain the steps in the process: from knowledge to model. I will end with a discussion about the challenges of complex knowledge-driven models, and the systematic errors that can occur due to questioning and extracting knowledge. All examples are created using the python library bnlearn.
When we talk about knowledge, it is not solely descriptive knowledge such as facts. Knowledge is also a familiarity, awareness, or understanding of someone or something, procedural knowledge (skills), or acquaintance knowledge [1].
Whatever knowledge you have, or want to use, it needs to be presented in a computer interpretable manner if you want to build a computer aided knowledge model.
This means that you need to design a system that is built on a sequence of process stages. Or in other words, a sequence goes from the output of the process into the input in the next process. Multiple simple sequences can then be combined into a complex system. We can represent such a system as a graph with nodes and edges. Each node corresponds to a variable and each edge represents conditional dependencies between pairs of variables. In this manner, we can define a model based on the expert’s knowledge, and the best way to do that is with Bayesian models. To answer the question, ‘Can we get experts knowledge into models?’ Well, it depends on how accurate you can represent the knowledge as a graph and how precise you can glue it together by probability theorem a.k.a. Bayesian graphical models. Still, there are some restrictions though.
The use of machine learning techniques has become a standard toolkit to obtain useful insights and make predictions in many domains. However, many of the models are data-driven, which means that data is required to learn a model. Incorporating expert’s knowledge in data-driven models is not possible or straightforward to do. However, a branch of machine learning is Bayesian graphical models (a.k.a. Bayesian networks, Bayesian belief networks, Bayes Net, causal probabilistic networks, and Influence diagrams), which can be used to incorporate experts knowledge into models and make inferences. See below some bullet points with the advantages of Bayesian graphical models, which I will stress throughout this article.