The industry as a whole is beginning to realize the intimate connection between Artificial Intelligence and its less heralded, yet equally viable, knowledge foundation. The increasing prominence of knowledge graphs in almost any form of analytics—from conventional Business Intelligence solutions to data science tools—suggests this fact, as does the growing interest in Neuro-Symbolic AI.
In most of these use cases, graphs are the framework for intelligently reasoning about business concepts with a comprehension exceeding that of mere machine learning. However, what many organizations still don’t realize is there’s an equally vital movement gaining traction around AI’s knowledge base that drastically improves its statistical learning prowess, making the latter far more effectual.
In these applications graphs aren’t simply providing an alternative form of AI to machine learning that naturally complements it. They supply the setting—the visual capabilities, dimensionality, and topology—for expanding the merit of the vectors at statistical AI’s core with a range of techniques including embedding, manifold learning, and clustering.
Utilizing AI’s knowledge base to better its statistical base via graph’s ability to increase machine learning’s aptitude is further testament to the undisputed truth that, for AI, “if you’re not using a diversity of approaches you’re limiting yourself in the generality of your system,” remarked Kyndi CEO Ryan Welsh.
The chief advantage knowledge graphs provide for machine learning is a relationship-savvy environment for depicting all the intricacies of the connections between individual nodes of data. A technique known as embedding is particularly useful in this regard. According to Katana Graph CEO Keshav Pingali, “Embeddings are a way to find relationships between entities that are not always obvious if you look at the connectivity of the graph.” In some use cases, embeddings simplify—if not obviate—the need for otherwise time consuming feature engineering.
In almost all deployments they identify non-linear relationships between entities to inform query results and searches for attributes used to build machine learning models. “The beauty of embeddings is if you do them right, those seemingly unrelated nodes that are far away in a graph end up close together in a three dimensional space,” Pingali observed. Such results are critical for building machine learning models with the proper weights and measures to maximize the use of training data for the most accurate models possible.
Embeddings are also a means of accelerating cognitive search capabilities to return nuanced results to users in a fraction of the time otherwise required to do so. Pingali mentioned there are multiple ways to perform embeddings, some of which involve Convolutional Neural Networks and web to web techniques. “What you do in a vector space model is you take every node and you map it into some point in a large dimensional space,” Pingali revealed. “And then the way that mapping happens is that there are related nodes, nodes that are similar in their properties, that may not be related together in the graph.”
Embeddings position those nodes together so the topology displays their similarities according to business problems. In the pharmaceutical space this approach is helpful for “hypothesis generation,” Pingali explained. “It’s very expensive to try out drugs in the lab to see whether they work, so if you can narrow down the candidates by using graph and AI, then that saves [companies] a tremendous amount of money.”
In other use cases, firms can leverage embeddings to search for the most relevant pharmaceutical elements or compounds to treat a healthcare issue like diabetes 2, for example.
