Answering the Question Why: Explainable AI

The statistical branch of artificial intelligence has enamored organizations across industries, spurred an immense amount of capital dedicated to its technologies, and entranced numerous media outlets for the past couple years. All of this attention, however, will ultimately prove unwarranted unless organizations, data scientists, and various vendors can answer one simple question: can they provide explainable AI?

Although the ability to explain the results of machine learning models—and produce consistent results from them—has never been easy, a number of emergent techniques have recently appeared to open the proverbial ‘black box’ rendering these models so difficult to explain.

One of the most useful involves modeling real world events with the adaptive schema of knowledge graphs and, via machine learning, gleaning whether they’re related and how frequently they take place together.

When the knowledge graph environment becomes endowed with an additional temporal dimension that organizations can traverse forwards and backwards with dynamic visualizations, they can understand what actually triggered these events, how one affected others, and the critical aspect of causation necessary for explainable AI.

Investments in AI may well hinge upon such visual methods for demonstrating causation between events analyzed by machine learning.

Correlation Isn’t Causation

As Judea Pearl’s renowned The Book of Why affirms, one of the cardinal statistical concepts upon which machine learning is based is that correlation isn’t tantamount to causation.

Part of the pressing need for explainable AI today is that in the zeal to operationalize these technologies, many users are mistaking correlation for causation—which is perhaps understandable because aspects of correlation can prove useful for determining causation. In ascending order of importance, an abridged hierarchy of statistical concepts contributing to explainable AI involves:

• Co-occurrence:This basic machine learning precept indicates how often certain events occur together. For example, machine learning results might show that peanut-allergy symptoms have a high co-occurrence with asthma or other health conditions.
• Correlation:Partially influenced by co-occurrence, correlation predominantly means there is a relationship between events. Significantly, it doesn’t denote what that relationship is.
• Causation:This concept is essential to explainable AI in that it illustrates why events occurred, or what caused them. For instance, findings might show that web page color, rather than product placement, is causative for upselling e-commerce customers.

Causation is the foundation of explainable AI. It enables organizations to understand that when given X, they can predict the likelihood of Y. In aircraft repairs, for example, causation between events might empower organizations to know that when a specific part in an engine fails, there’s a greater probability for having to replace cooling system infrastructure.

Causation in Time

There’s an undeniable temporal element of causation readily illustrated in knowledge graphs so when depicting real world events, organizations can ascertain which took place first and how it might have affected others.

This added temporal dimension is critical in establishing causation between events, such as patients having both HIV and bi-polar disorder. In this domain, deep neural networks and other black box machine learning approaches can pinpoint any number of interesting patterns, such as the fact that there’s a high co-occurrence of these conditions in patients.

When modeling these events in graph settings alongside other relevant events—like what erratic decisions individual bi-polar patients made relating to their sexual or substance abuse activities—they might differentiate various aspects of correlation. However, the ability to dynamically visualize the sequence of those events to see which took place before what and how that contributed to other events is indispensable to finding causation.

Explainability and Accuracy

The flexibility of knowledge graph schema enables organizations to specify the start and end time of events. When leveraging speech recognition technologies in contact centers for sales opportunities, organizations can model when agents mentioned certain sales products, how long they talked about them, and the same information for customers. Visual graph mechanisms can depict these events sequentially, so organizations can see which led to what. Without this temporal method, organizations can leverage machine learning to specify co-occurrence and correlation between products.

Nevertheless, the ability to traverse these events at various points in time allows them to see which products, services, or customer prototypes generate interest in other offerings. This causation is determinate for increasing the accuracy of machine learning predictions about how to boost sales with this information. As valuable as this capacity is, the more meritorious quality of such causation is that the explanation for these predictions is not only perfectly clear, but also able to be visualized.

Explaining AI

Causation is the basis for understanding the predictions of machine learning models. Knowledge graphs have visualizations enabling organizations to go back and forth in time to see which events are causative to others. This capability is vital to solving the issue of explainable AI.

About the author: Jans Aasman is a Ph.D. psychologist, expert in cognitive science and CEO of Franz Inc., an early innovator in Artificial Intelligence and leading provider of semantic database technology and knowledge graph solutions. As both a scientist and CEO, Dr. Aasman continues to break ground in the areas of artificial intelligence and knowledge graphs as he works hand-in-hand with numerous Fortune 500 organizations as well as U.S. and foreign governments.