In this blog post, I will talk about an important concept, which is often overlooked in data mining and machine learning: explainability.
To discuss this topic, it is necessary to first remember what is the goal of data mining and machine learning. The goal of data mining is to extract models, knowledge, or patterns from data that can help to understand the data and make predictions. There are various types of data mining techniques such as clustering, pattern mining, classification, and outlier detection. The goal of machine learning is more general. It is to build software that can automatically learn to do some tasks. For example, a program can be trained to recognize handwritten characters, play chess, or to explore a virtual world. Generally, data mining can be viewed as a field of research that is overlapping with machine learning and statistics.
Machine learning and data mining techniques can be unsupervised (do not require labelled data to learn models or extract patterns from data) or supervised (labelled data is needed).
In general, the outcome of data mining or machine learning can be evaluated to determine if something useful is obtained by applying these techniques. For example, a handwritten character recognition model may be evaluated in terms of its accuracy (number of characters correctly identified divided by the number of characters to be recognized) or using other measures. By using evaluation measures, a model can be fine-tuned or several models can be compared to choose the best one.
In data mining and machine learning, several techniques work as black-boxes. A black box model can be said to be a software module that takes an input and produces an output but does not let the user understand the process that was applied to obtain the output.
Some examples of blackbox models are neural networks. Several neural networks may provide a very high accuracy for tasks such as face recognition but will not let the user easily understand how the model makes predictions. This is not true for all models, but as neural networks become more complex, it becomes more and more difficult to understand them. The opposite is glassbox models, which let the user understand the process used to generate an output. An example of glass box models are decision trees. If a decision tree is not too big, it can be easy to understand how it makes its predictions. Although such models may yield a lower accuracy than some blackbox models, glassbox models are easily understood by humans. In data mining, another example of explainable models are patterns extracted by pattern mining algorithms.
A glassbox model is thus said to be explainable. Explanability means that a model or knowledge extracted by data mining or machine learning can be understood by humans. In many real world applications, explanability is important. For example, a marketing expert may want to apply data mining techniques on customer data to understand the behavior of customers. Then, he may use the learned knowledge to take some marketing decisions or to design a new product. Another example is when data mining techniques are used in a criminal case. If a model predicts that someone is the author of an anonymous text containing threats, then it may be required to explain how this prediction was made to be able to use this model as an evidence in a court.
On the other hand, there are also several applications where explanability is not important. For example, a software program that do face recognition can be very useful even though how it works may not be easily understandable.
Nowadays, many data mining or machine learning models are not explainable. There is thus an important research opportunity to build explainable models. If we build explainable models, a user can participate in the decision process of machines and learn from the obtained models. On the other hand, if a model is not explainable, a user may be left out of the decision process. This thus raises the question of whether machines should be trusted to make decisions without human intervention?
In this blog post, I have described the concept of explainability. What is your opinion about it? You can share your opinion in the comment section below.