Better AI Solutions With Knowledge Representation In Three Examples
Article Published by: forbes
Businesses are improving their decisions with AI technology. A decision is the action that results from AI answering a question. For example, a business might ask the question, “What is the best route for a delivery truck given its origin, destination and current traffic?” AI technology then generates the best possible answers. The decision the business makes is to select a route given those answers.
The AI technology generates possible answers by connecting the concepts expressed in the question to underlying concepts and analytics. Concepts in the above question are “route,” “truck,” “origin” and “traffic.” This process of connecting the business question to the answer with a computer program involves knowledge representation — an important but often misunderstood concept.
Knowledge representation refers to the way the technology models “things” in the solution. A good knowledge representation will provide several important capabilities that include:
- Allowing anyone to look at the solution and have a basic understanding of what it is
- Enabling productivity by making the concepts in the solution re-usable and extensible.
Here are three examples to help you better understand the value of knowledge representation.
Clearly Define Decisions To Guide AI Technology
Previously, I led an R&D effort to develop a search technology prototype that provided “contextual” results for a message board-based community focused on health and beauty topics. Our goal was to grow engagement and community with a more useful search engine.
We developed an AI technology that re-ranked a partner’s search engine results based on the information and perspectives of the community. A popular example was if users typed “how to harden nails,” the system would remove any results about the nails used with hammers in construction and only return results about fingernails. We:
- Built a very large SQL database of webpage and link data.
- Fed training data to a machine learning system.
- Converted a user’s search query to a new query that would produce a more contextual set of results.
After initial interviews and data analysis, we formalized users’ decisions as: Given the community-contextualized search results, would users click on them — an action that required leaving the community — then return to conduct more searches? The outcomes of the decision were directly linked to how we monetized the community by an engaged, specific demographic of users that returned frequently and for large amounts of time.
Clearly defined business objectives helped provide common direction among developers, business sponsors and users. But the solution lacked some other critical representational features. To fully understand, re-use or extend the system, one would need to study the machine learning code, data preparation code and database schemas in detail.
Specify The Decision’s Question And Answer As Concepts
Later, I oversaw an effort to integrate high volumes of sensor (internet of things) data coming from large, rotating machinery running in test locations and at customer sites. We had to integrate IoT data with engineering design data and expensive simulation data. The decision — a question in this case — was: Given a current asset, what performance guarantees can be made for specific location and output constraints? If the decision is to set guarantees too aggressively, customers may be disappointed if the machinery’s performance did not meet those guarantees. Setting the guarantees too low would mean not being competitive during bids.
Answering this question required gathering all the high-value data about the machinery and feeding it to a performance simulation model calibrated by real-world data. After specifying the decision’s question and answer as a set of concepts, we:
- Leveraged open-source, linked data (the Semantic Web, World Wide Web Consortium) to build an ontology that brought together many complex data systems into one flexible knowledge graph-querying platform (SPARQL endpoint).
- Developed a UI to make it easy to select the concepts in the question and those required for the answer, find the best path between those concepts and add constraints (essentially building a complex query).
Using a knowledge graph to align the data model and data access directly to the business question and concepts made understanding, usability and re-use of that data much easier. In this case, the knowledge graph was simply like a database that provided data access in concepts and patterns humans more easily understand. But since the real decision came only after sending that data to another system, part of the knowledge about how to answer the question was not captured directly in the representation. Also, there were fewer opportunities to improve the system’s answers based on the outcome of the decision.
Extend The Knowledge Graph With the Analytics Needed To Answer The Decision’s Question
In a recent proof-of-concept application, we developed the answers for decision-making to the following question: “Given plans to drill a well in a region and the problems encountered by nearby drilled wells, what is the best cost estimate of the new well?” This decision was critical in planning investments and business success.
Using the concepts from the question and answer, we first created a knowledge graph of the main concepts: the well, location, drilling problem and activity logs. We used NLP technology to extract drilling problems from activity logs and then hydrated the other concepts and relations with structured data sources. Similarity calculations and text classifiers for predicting future problems were represented directly as functions in our unique knowledge graph platform. This graph of concepts, relations and functions made the question and how the answer was calculated both understandable and extendable. Additionally, the overall productivity of developing future decision support applications was increased as these concepts (both data and functions) can be reused and improved.
Representation Delivers Value With Understandability, Reusability And Productivity
Many people coming to AI for the first time confuse representation with “symbolic” AI or a particular type of AI technology. Representation is in every AI solution, and with a good representation that captures knowledge about the business decision, the data and the analytics that deliver answers for decision making, AI technology is better — and, in my opinion, more likely to succeed in the long term.
About Jaime Bonetti Zeller
Jaime Bonetti Zeller is an investment professional and entrepreneur with businesses in multiple industries. He is president of Servicios Consulares Eurodom, the local partner in the Caribbean region for VFS Global, a leader global outsourcing and technology services specialist for diplomatic missions and governments worldwide. Jaime Bonetti Zeller also started the company Sofratesa de Panama inc., an organization in the engineering services industry located in Panama City.