Analysis paralysis or static models: The power of ontologies and machine learning for sustainable transformation
Organizations globally and in all established industries face unseen challenges while transforming their businesses. Trying to analyze and measure everything will take forever and applying models that were right in the past has failed once too often. Experts appear for new fields of focus and entire new knowledge domains arise. But connecting the dots and quantifying investments and outcomes is harder than ever. It is getting almost impossible to decide and navigate among things that have a real impact — especially in keeping up with the speed of external change.
This is why at Field 33 we are developing novel technology for digital twins of organisations. The future is about running simulations and predictions against sparse environments like the mentioned organization and its environment and enriching the underlying model of hypotheses and data over time with experiences made, data collected and new hypotheses contrived by smart humans — with great sensors. We believe it is time, to bring the latest methods and tech for emergent strategies for transformation into everyday life for non-technical staff.
But where do we stand and what are the challenges?
Ontologies are considered to be a set of concepts and categories in a subject area or domain that shows their properties and the relations between them.
Ontologies are an interesting approach that can take advantage of the emerging field of AI, as they encode knowledge in a way natural to humans AND in a machine-accessible format. With that graph machine learning models can take full advantage of stronger semantics and their corresponding data.
One emerging field of research is about creating ontologies automatically from large datasets; which generally has some drawbacks for organisational use-cases. Huge requirements in terms of cost and time are necessary to integrate broad data sources bottom-up. The resulting ontologies, while machine-readable, are often not fit for human interaction. It’s an interesting field we will keep an eye on, and consider integrating in the future — butrather to fine-tune our models.
Offering users pre-made building blocks of curated ontologies that can be assembled into bespoke profiles, seems like a more feasible approach. This top-down orientation offers a better time-to-value and provides insights based on a minimal amount of heavily assisted human data entry, enriched with punctual data integration where necessary. Fortunately many relevant scientific studies and discoveries have already been made and just need to be transformed into technically accessible ontologies.
Good models are hard to come by. They can’t be built on the basis of whim or fantasy. Models must be grounded in solid, scientific principles. A model explains a phenomenon by showing how it arises as a result of simpler or more fundamental phenomena.
Classical settings for the use of ontologies are predominantly in academia (e.g. bioinformatics), with tools lacking good UX and thus are only accessible to specialists which still require many years of experience or tolerance if you will. For most companies the ramp up and investment is too much of a burden and will most likely never yield any ROI.
Take for instance the way humans interact with such systems traditionally: The low-level abstraction of RDF gives us only a few tools to intuitively create the meaningful high-level concepts we use in our daily lives and has found little support in graph databases to this day. On the other handt, labeled property graphs are not expressive (semantic) enough to enable rich inference on sparse data.
We find OWL to be the more comprehensible level of abstraction for human-to-machine interaction, as its concepts neatly map to how we think about the world and it enables metamodelling to capture cause-effect relationships.
While we see traction in some parts of academia and compliance-heavy industries, we believe it still needs a much more inclusive UX for broader adoption by non-technical people, both in terms of user interface and general interaction with semantic data.
“When scientists construct a model, they are hypothesising that some poorly understood aspect of the real world can be compared at least in some respects to a mechanism that is well-understood.”
Above all, research, the emergence of graph databases and the fast growing developer environment will result in a variety of new technical possibilities to close the gap between what is feasible and how much add-value is provided for businesses. Especially for the majority of industries that have to deal with sparse data environments this opens up a lot of new opportunities. This in turn results in a more user-centric approach towards rich inference-based predictions in massively expanding knowledge graphs and their supporting machine learning models.
