In our previous blog post, we have introduced an essential issue for companies which aim to create effective ML-based products. Namely, what process and actions are required to implement Artificial Intelligence solutions which concretely answer market’s and customers’ needs?
Well, many things change when Machine Learning models need to be applied to business needs and not developed within the protected environment of academia. Starting from the latter, currently we can identify substantially two steps that are iterated for the implementation of a ML-based solution, namely the model development phase and the following evaluation offline stage, which is testing the model itself using a predefined dataset. This process is circular, as it involves the repetition of evaluation and model fixes until the accuracy of the latter reaches a satisfactory level. And it actually works when the procedure is completed within the academia safe environment. But what happens when we switch to the market scenario?
Obviously the situation is very different in the business world. When we have used the term safe, we have referred mostly to the simplified nature of the context in which ML academia researchers work, specifically in respect to their starting datasets and model evaluations. But, when we talk about the development of a real ML-based product, which needs to face real business problems and customers’ needs, some other steps are required in addition to the core activity, which includes the model creation and evaluation. At Aptus.AI we have been focusing on this issue in recent years, therefore reaching a more consistent procedure, which includes three further phases:
It is clear that the first two steps need to be completed before the model designing, while the third phase is an additional model evaluation, which takes into account the change of data and information on which the Machine Learning system needs to work.
Many business problems are difficult to be solved even if they are very clear and visible. Things get harder when issues are more complex and multifaceted. But the situation becomes even unmanageable if problems are not defined before starting a work. This is why at Aptus.AI we give the greatest importance to the problem definition phase, which includes meetings and confrontations between managers, technicians and customers. This step is also required to set the following phase, which is the preparation of a specific dataset - which we will discuss in our next post. The problem definition needs to take into account the market situation and the customer’s needs, but also to consider all the technical limits and possibilities in respect to ML technologies. In this respect, it is also essential to set a definition of success, which is exactly a definition of the business goals related to the development of a specific ML-based solution.
Turning back for a moment to the academia world, we can say that, in that context, usually success is identified with models accuracy, but - as we said in our previous post - accuracy can be considered an informative proxy about an AI product, not directly proportional to a business-centered success. It is not always true that accuracy is the right metrics to evaluate a Machine Learning system. This is why at Aptus.AI we focus on how to model the problem we want to solve. To make it clear, we can quote Daitomic, our interactive AI platform created for financial compliance management. This software automatically detects and distinguishes banking obligations and penalties, but it also considers - at the same time - the major criticism of the latter, because - also in this case - the overall accuracy of the ML model is important, but it needs to go together with the awareness of what is actually useful for users - that is, for customers. This one is the only approach to follow to create an actually effective AI solution. Just like Daitomic.