One of our goals while creating the Akbank Lab Medium account was to share both advantages and difficulties of what we have experienced in our projects that we are working on. As Akbank innovation center, we both take part in innovative projects with our relevant colleagues within the bank, and aim to create common added value by bringing them with the innovation ecosystem together.
On the Akbank Lab side, the main topics we focus on are advanced analytical studies. Akbank definitely supports all its work with advanced analytical applications by working with its internal resources and external business partners. The reflection of this on the stakeholders can be in different ways, such as providing personalized suggestions, being with the customer in case of need, and providing faster services. In this sense, the steps of all processes which have both internal and external customer touchpoints are supported by advanced analytical applications. In this article, we wanted to share with you information about our "Automated Machine Learning" platform that we have recently implemented.
Although it has been 60 years since the emergence of the machine learning concept, we started to feel its effects strongly both on our daily and business lives especially after the 2000's. With the transformation of this concept into an applicable technology, the development of competencies in this field and the increase in accessibility of this technology, it is not difficult to say that the application and its area of influence are expanding every year, and this technology will develop further and affect us much more in the future.
Machine learning has had a huge impact on the way of doing business and will continue to do so. The ability and speed of this technology to analyze huge amount of data motivates institutions to invest in this technology in almost every field. As the usage areas of machine learning became widespread and started to be prioritized by more institutions, new tools and methods started to be produced. In addition, following this constantly evolving technology and having the most up-to-date knowledge became quite challenging even for the strongest institutions. At this point, the importance of "Automated Machine Learning" platforms which support faster and higher performance in developing analytical models has increased considerably in recent years.
As the institutions which have wide customer base, strong technological infrastructure and huge amount of data, of course the banks are most eager to adapt this technology. In this journey, which started years ago with corporate analytics platforms, Akbank has been placing great emphasis on improving its analytical competencies for many years as an institution that uses open source programming languages such as Python and R, and then machine learning and automated machine learning methods. As Akbank Lab, we are trying to contribute to this journey with different projects and to take advantage of cooperation opportunities with third party players who are able to give speed and competence in the innovation ecosystem.
We carried out one of these projects together with our Analytics team, and we aimed to test an automated machine learning platform mentioned above to transform our model development cycle into a faster and more efficient level. During this project, which lasted for about 1.5 months, we chose H2O.ai from Silicon Valley as our business partner and we put this platform into use at the beginning of 2020. You can see that H2O.ai is one of the most innovative companies in this field and cooperates with many large institutions on a global scale, also stands out in Gartner research. So, what are automated machine learning platforms and how do they add value to organizations?
Figure 1. Data Science and Machine Learning Platforms 
Automated Machine Learning Platforms
We can base the steps in Figure 2 to summarize the basic steps followed by organizations in different sectors for model development cycles. In these steps, there are development areas to be simplified in order to achieve speed and accuracy (precision). For example:
• Existence of a large number of models produced and updated periodically
• Time-consuming model development cycle
• Explainability of models developed with nonlinear algorithms
• Difficulty in creating documentation for models
• Access to competent data scientists and need to constantly keep these people updated on different types of models (e.g. natural language processing)
Automation in modeling & profiling, testing & dissemination and documentation steps, together with the increase in data production and need to personalized customer services, provides significant efficiency today. With these platforms, the duration of these steps, which normally take a few months, can be reduced to weeks. Thus, a faster model development and testing opportunity can be created.
While we were working on our possible focus areas with the Akbank Analytics team to make our advanced analytical work more innovative, we learned that there are more than 100 models created and managed by this team. We aimed to increase the efficiency and to improve the customer experience by new analytical studies developed much faster with such an automation project.
Figure 2. Key Steps in the Model Development Process
We chose H2O.ai as our business partner in this project. With automated machine learning, it is possible to shorten model development processes, train models faster, and develop more accurate, precise and “explainable” models. We have seen that some of the features of the H2O.ai platform also provided some significant gains.
• Easily understandable printouts / automatic documentation even for non-data scientists
• Automatic variable generation and model development
• Ready-made model library specific to different usage scenarios which is created and kept up-to-date by Kaggle Grandmaster  level data scientists
• Fast analysis of variables with a user-friendly interface (Figure 3)
Figure 3. H2O.ai Platform Sample Interfaces
As Akbank Lab, in the business areas we have determined jointly with our business unit we propose such innovative solutions, which we want to try and see the results of, to the Innovation Selection Committee in order to take approval to test them. We initiate the PoC (proof of concept) process for approved solutions and then with our business unit and the Innovation Committee we evaluate how close we are to the targeted results after PoC. In this process, we experience an environment that nurtures the culture of innovation as well as improving better business results by working with a Fintech, which we have determined as our business partner with Akbank teams.
Our Project Process
In order to test the platform of H2O.ai and determine its possible benefits for us, we have chosen 5 most up-to-date and advanced models together with our Analytics team in line with the PoC approach to see whether there has been an improvement in both the model development times and the success rates of these models. Also, we prepared the server and data requirements for 5 models to install the platform in bank systems for production. Because the technical compatibility of this platform with our systems and understanding the requirements for optimum performance was also an important point.
When we reconstructed these 5 up-to-date models with this platform within the scope of the project, which lasted about 1.5 months, we saw that the accuracy in most of the models created by this platform was better. Therefore, considering our entire portfolio of models, including less up-to-date models and algorithms, we came to the conclusion that the potential contribution of such a platform could reach significant levels. While testing the improvement in our models, we compared the models created within the platform with the past performances of these models to calculate how much additional benefit Akbank would have if we had used it in the past.
In addition, we have seen that many steps involved in the model development process can have significant gain in time, and a 2–3-month process can be reduced to 3–4 weeks. As a result, we realized that we could achieve our goals determined at the beginning of the project and decided to add this platform to our bank.
With the launch of the platform at the beginning of 2020, we started working on the adaptation of the platform in our monthly modeling cycle with our Information Technologies Infrastructure teams. While this process was continuing, we continued model development studies in line with our 2020 planning. We have completed more than 30 model development studies and tests in the first six months as we aimed while giving buy decision of the platform. We aim to complete the development processes for more than 40 models in 2020, too. Thanks to this platform, we quickly renewed a large number of comprehensive models and had the opportunity to meet the needs of business units in new model making by many trials in a short time.
 https://www.forbes.com/sites/janakirammsv/2020/02/20/gartners-2020-magic-quadrant-for-data-science-and-machine-learning-platforms-has-many-surprises / # 37f840813f55
 Data scientists who reached the top title as a result of the activities like competitions, problem solving, etc. on the Kaggle platform, which created is created by Google and full with data scientists and machine learning experts