What is XAI? How to Apply XAI
If you believe the reports, AI is simply uploading big data, running an AI algorithm, and getting results. Unfortunately, it is not that simple. Practical methods to apply AI are often limited to guidelines on choosing the right AI algorithm. As a result, AI solutions are black boxes, with unclear scope, that are not trusted by the workforce and not integrated into the organization's strategy cycle.
This article provides practical guidelines to create a sustainable AI solution that is integrated into the organization's strategy development cycle and uses explanations to guide humans in a plan-do-check-act process. The result is an XAI solution combining symbolic reasoning — business rules — to deal with common-sense knowledge and machine learning — statistics — to deal with uncertainty. In AI terms this is named a 'hybrid solution'.
Any easy-to-remember six-step process merits some warnings. This is a generalized method, and blindly following a generalized method as a recipe on a specific case never yields good results. Understanding the underlying knowledge and thinking about how to best apply it in a specific situation will. So, when applying this method, make sure to understand the motivation behind each step and assess how to apply each step to your specific case. Furthermore, the steps need not be applied linearly; instead they may be revisited as often as needed.
Step 1: Concept Model Definition
It all starts by agreeing, as a business, on the meaning of the important concepts in your business. A business that can't state clearly who they target and what it means to be successful is unlikely to be efficient, effective, and successful in the application of AI and decision-support systems.
For most businesses, a returning customer is an important concept. If we use AI to predict the likelihood of a new customer becoming a returning customer, how do we define clearly what a returning customer is? A well-defined business concept should be SMART (Specific, Measurable, Achievable, Realistic, and Timebound). Is it a customer who ordered within two months after their first order and placed the last order within the past 12 months? This example demonstrates that SMART definitions form the basis for a good scope definition, result in concepts with a clear and distinct meaning, and relate a concept to other important concepts.
In my experience, businesses that have well-defined concepts are better at communicating with their customers, partners, and technology departments and, therefore, are more successful in finishing technology projects, including AI innovations.
Step 2: Task Definition
To collect the right data, you have to understand the task that you want to support. I recommend that you always start with a basic statistical analysis on the available data. This opens the conversation and discussion in the organization and will lead to better understood and defined concepts.
In our returning customer example, a distribution of orders in time may confirm that, indeed, when customers wait more than 3 months to place a second order it is unlikely that they will return. But what predicts that they will return when they place a first order?
The business is likely to have some good ideas and some preconceptions, which need to be brainstormed, analyzed, and tested. Are multiple outcomes acceptable, or do you need just one outcome? Is the result of the task a classification, a ranking, or a number? Agreeing on these questions is important for step 4.
Step 3: Data Collection
The good news about this decade is that most organizations are already collecting data. Are we collecting the right data for our objectives? Most of the time, data is collected for administrative purposes and not for predictive purposes. For predictions we may need to collect more data or get new data sources.
Let's use the returning customer example again. Maybe the parking conditions at the time of the first order are an important predictive factor in your business for a customer to make the decision to return or not. Did we collect that data? Most likely we did not. Can we get it? Most likely we can.
Step 4: Technique Selection
Now we can choose a suitable AI technique because we know from the previous steps what kind of data we have as input for our decision support task, how the data elements relate, what kind of decision-outcome we accept, and what we already know about the task that needs to be supported.
You may have heard about 'random forest' and 'deep learning'. These are two of many AI techniques to choose from. Which technique to apply depends on the characteristics of the task and data. There will be another article about these techniques, but I can already tell you that it is definitely not the other way around: it is not the technique that tells us what to do! No, it is the task that dictates what technique to use. So, don't use a hammer when you need a screwdriver!
Step 5: Explanation Generation
The major concern with decision support systems in general is that if users do not trust the underlying model or a prediction they will not use it. It has been proven that a good explanation may increase the trust in the system. However, if the explanation is complicated, nonsensical, or difficult to understand, the distrust of the system's decision will increase. So, generating a good explanation is a separate step — and an important step.
AI experts tend to claim that models that are good in predicting are not good in explaining and vice versa. This stance suggests that one has to choose between transparency and accuracy. I don't agree. To the contrary I claim that high accuracy and explainability can, and should, go hand-in-hand.
Generating the explanation does involve a second model. The explanation model has to receive information from the predictive model, get the distinguishing factors, select the factors that contribute to the user's confidence, and present those in the right way.
Suppose that, in our returning customer example, we identified that a customer is not likely to return because of the lack of parking spaces and the acquired product is not in the top 10 for returning customers. However, the customer asked a question and we know this increases the likelihood that a customer may return. Then, it makes sense to provide an explanation to the user that includes both evidence for and against the conclusion. This makes the system trustworthy and intelligible for users, and this has proven to increase their trust in the advice of the system.
Step 6: System Evolution
Finally, there needs to be a way to track and analyze errors and new trends in order to evolve the system. Besides tracking the accuracy of the predictions automatically, we need to actively look for side effects, biases, and new knowledge. That is, WE — being creative and experienced business people — must be supported by a good automated workflow process to improve the decision-making process.
In our returning customer example, each customer that returns but was not 'flagged' by the system as a 'returner' should be analyzed. Do we know something that the system did not know? For example, the customer returns every spring and only buys something for the garden; if it makes sense, it should result in an update of the system's knowledge or system's scope description.
Better explanations should result in increased trust and also in increased performance. These are the KPIs of our XAI solution and they should be measured.
The 6 steps of XAI versus AI practices today
Maybe you expected steps like: write use cases, create MVP, integrate MVP into IT infrastructure, refine the business process, assess performance, automate feedback loop. This is the typical terminology used in the AI world today, where the focus is on getting fast results, often part of a "fail fast, learn fast" culture. As a consequence, some of these efforts are not sustainable and will not result in business value. AI adoption has tripled in 2018, moving AI towards the Gartner-hype-cycle peak. Now that AI is becoming increasingly mainstream, more conservative companies have good reasons to enter this arena. These companies will ask for a method that results in more reliable project outcomes and integrated business systems.
We don't want decision support systems that result in the headline: "AI model of XYZ uses last name of applicant to determine insurance eligibility; is that legal?"
AI solutions assume that big data includes knowledge about the correct input-output relationship. However, what this knowledge is about remains a black box to us and therefore the machine-made decision maker is comparable to an oracle. In XAI we still use these AI techniques but add an extra diagnostic feedback mechanism.
This is the fifth article in this series. As an expert in decision support systems development, I have been promoting transparency and self-explanatory systems to close the plan- do-check-act cycle. All too often I come across modern systems that have similar issues, and face the same fate, as the legacy systems they replaced because the domain experts or end-users are not involved in the feedback loop. My impression is that the journey is starting all over again as organizations begin using AI technology as black box systems, but that is not necessary. This is one of my contributions to a topic that is close to my heart. Next time we will describe what makes an explanation a good explanation.
Let me know if you liked this post by sharing it in your network.
Acknowledgements: I am grateful to Patricia Henao for helping me by editing the articles, providing her reflections, and creating supporting training material.
# # #