How ontologies can give machine learning a competitive edge

Ontology is a concept with slightly different meanings in the contexts of philosophy and information science that is probably unfamiliar to many. In philosophy, it refers to the study of being. In information science, it describes a set of concepts and categories in a subject area or domain that shows their properties and the relationships between them.

In practical business terms, ontology is the architecture that connects numerous sources of information. It builds relationships, qualifies and interconnects data from multiple domains by automatically tagging and categorising that data. You can think of it as a means of communicating and resolving semantic and organisational differences between databases in order to enhance their integration.

The ability to work with unstructured, semi-structured or structured data formats means ontologies can connect and qualify data without any need for standardisation. They streamline the process of identifying core concepts, improving classification results to collate critical information. As a result, data can be found and analysed faster.

For engineers, ontologies are a flashlight that helps them find the nuggets of treasure hidden at the bottom of an ever-rising sea of data.

That rising sea is becoming a growing problem for many companies. And while machine learning (ML) and deep learning have enabled enterprises to glean insights from their data and drive all sorts of efficiencies, we are now approaching a data ceiling that could block further progress. Businesses are discovering that too much data can be overwhelming to analyse, resulting in value-destroying complexity, time and cost. Some forward-looking enterprises, however, are using ontologies to stay on top of ballooning datasets and get more from their investments.

According to conventional wisdom, the more data you have, the better machine learning and artificial intelligence in their various guises can work some kind of magic. This is misleading. By some estimates, as much as 85 per cent of AI projects fail to meet business objectives. The reason stems from a lack of understanding about how to utilise large volumes of data. In reality, AI projects frequently fail because of problems with the architecture of the information rather than the quantity. Ontologies can make a big difference.

While ontologies have been used by diverse organisations in a range of industries, some enterprises have struggled to deploy them. In some instances, there has been a lack of awareness and understanding of how to harness their full potential and even a shortage of expertise.

The implementation of ontologies is a process of continuous improvement that can be rolled out in stages rather than enterprise-wide, so it doesn’t have to be an all-or-nothing proposition. One way to gain comfort with them is to start small with a few trials and then broaden to larger data sets. Even if you discover an error, you can still fine-tune the ontologies without jeopardizing the entire AI and ontologies initiative.

There are some practical things that engineers can do to develop and adopt ontologies.

First, identify the ontology’s purpose. For example, it may be quality control, fraud detection or customer authentication. Formalise the processes within the enterprise for this use case. Using the SIPOC process improvement method to evaluate supplies, inputs, processes, outputs and customers can help identify any problem areas across different inputs and outputs.

It’s also wise to adapt and validate the ontologies as they are trialled. Identify the root causes of any problems, technical requirements, workflows or decision processes.

Finally, create and test the ontologies to assess how efficiently each one performs. Adjust them as necessary - they need to be agile and able to solve the bigger data-architecture challenges that can thwart AI projects.

According to market research firm IDC, by 2025 the world could generate more than 175 zettabytes of data. To put that into context, one zettabyte is a trillion gigabytes. Given how increasingly reliant businesses are on data to make informed decisions, there is now a pressing need for engineers to incorporate ontologies into their data science initiatives and put this rising tide of data to good use.

A common mistake that AI specialists and data analysts make is to view ontologies as purely a tool to structure data, like a taxonomy. Ontologies can do much more than that. They can mimic logical reasoning, and thanks to their ability to automatically tag, categorise and link information they streamline the AI training process. They enrich datasets and provide faster analysis. If AI is the vehicle that relies on algorithms to power its engines and propel itself forward, ontologies are the fuel. To maximise the value of your data and avoid getting overwhelmed by it, give ontologies a try.

Pascal Feillard is automotive senior solution director at Capgemini Engineering