A well-set task means 30% of process optimization success, according to Eugene Kolesnikov, the Head of the Big data and machine learning focus area at Jet Infosystems. We talked to him and discovered how to prepare an online shop for the Black Friday and who has a better understanding of big data monetization.
Interviewer: AI Conference (AI)
Respondent: Eugene Kolesnikov (E. K.)
AI: What business tasks can big data and machine learning solve?
E. K.: In general, machine learning (ML) and big data analytics allow to solve any optimization problems related to these data. For instance, talking about manufacturing environment, these tools allow to improve operation performance using rolling mill yield data, such as its speed, output etc. In banks, big data analytics and machine learning help to significantly improve user experience, business processes, as well as fraud detection and security in general. In the retail sector, Big Data and ML bring customer relationship to a new level: marketing promotions, customer journey optimization, enhancement of user experience while working with a website or other company interaction channels.
ML allows to optimize (in other words to reduce expenses or increase revenue) all of these tasks by 5-10%, and sometimes by 30%, depending on a problem.
AI: What machine learning issues do specialists face?
E. K.: The key ML difficulties include the lack of data and the absence of well-set task. Currently, the market considers artificial intelligence something like Red Queen from the Resident Evil movie where the machine talks in a female voice and shoots with lasers. But that's not the case. ML always solves an accurate business task. The process requires a clearly established goal and data that can solve it. This aspect is most frequently troubled: there is a lack of data, business elements, or internal understanding.
The first and greatest complexity caused by a customer is the following: we need to find a person able to set a business task along with us. ML is a certain thinking style that can explain what my colleagues and I are focusing on. The first 30% of project success will be provided by choosing an appropriate task and a correct way to solve it.
AI: And what about technological issues?
E. K.: Nowadays, there are a lot of ML technological tools for solving certain business tasks. The problem is the absence of understanding how to live and work with them.
AI: When, and to whom, should this understanding come?
E. K.: It is a convergent process: on the one hand, it is encouraged by integrators (our colleagues and we), on the other hand, there is an internal need that will, sooner or later, make a company re-build the internal logic of understanding. You know, it is like the manufacturing automation during the first technological revolution: business processes have been changed from within. One had to think differently in order to automate them using IT. CRM and ERP were integrated in the same way.
ML is moving in the almost similar direction. Nobody doubts that the technology is useful and able to solve a variety of problems. The most important thing today is the internal understanding and customer adaptation to new technologies. We can see that major players are rearranging themselves, but it is quite hard. They are pioneers with incredibly huge volumes of data. For example, MegaFon has dedicated a whole company for this purpose – oneFactor*. Sberbank is building a team within its structure and growing in this area.
There are new companies aimed at technologies. However, they are initially technological and do not have to readjust anything in their operations.
A positive aspect is the fact that ML is a pretty young and debated tool, thus opening ML or Big Data subdivisions, companies aim to announce this news. It plays a significant role in the further popularization of tools.
AI: Which of the Jet Infosystems projects was the most interesting for you? Why?
E. K.: We participated in the M.Video hackathon, got a winning place, and consequently kept developing a single project together. It is designed to convert feedback about goods into clear characteristics.
For instance, let’s take a washing machine. It collects dozens of feedback like “well operates” but some of them are meaningful: “can be placed under a sink”, “narrow drum” etc. People describe important features that a company does not put into a product card. The task is to select such characteristics from the text and add them to the product card using automated tools.
AI: Which of the projects was the most complicated?
E. K.: The order from one of the top 10 banks. We had to detect internal fraud among operators based on their behavior in various systems. The customer provided us with data on employees’ behavior in certain systems. Using these data, we had to predict worker’s disposition to fraud.
It was huge amount of data and their sources that caused difficulties. We rebuilt inner understanding of detecting swindlers. It was quite complicated to cooperate with the customer within ML and to provide quality patterns of inspection results. The work on project was long and diligent.
AI: What are you working on at the moment?
E. K.: We are primarily working on manufacturing. Now, we are solving the majority of tasks, the most complex and interesting ones, for steelmaking companies. They have lots of data and the understanding of attracting benefits, monetizing ML and its methods. The 1-2% increase in a certain metric of the manufacturing sector brings revenue amounting to hundreds of millions per year.
Many retail projects, including tasks for marketing campaigns and optimization of internal processes related to turnover and product grids.
As to banks, it is various tasks for improving capacity, which will talk about, and security. Besides, marketing: one should realize what to offer customers and how.
Today, we can see a great demand that exceeds supply on the market.
AI: Eugene, you are going to tell AI Conference about the solution predicting a change in the IT resources load. Who will be interested in this topic?
E. K.: It will be interesting for everyone dealing with servers in their daily work, in other words for any company. The key question here is: what will happen if we will grow by 30% tomorrow? It leads to other ones: what systems will be loaded? Which of them should be expanded?
Unlike you and me, common users, a company is not able to attend the electronics market and buy an extra system unit. Machines are assembled due to customer’s demands and delivered in approximately three months. A plan should be elaborated half a year prior to the actual need. Therefore, our case study is quite significant.
We have methods allowing to predict the behavior of both certain machine and the whole system, depending on business features. For example, an online shop announces the Black Friday and increases the amount of customers threefold. What systems will suffer maximum load? What aspects should be taken into account: CPU, random access memory, or something else? What should be supported right now? If it refers to a virtual farm, one should increase their capacity to overcome the rush hour.
We are building two vectors, showing where the load will grow and explaining what will happen with the information system, depending on business performance spurts: the increase in the amount of cash desks, the development of shops, and support calls. All of this has a strong impact on IT systems, which affect company operations.
* One of the company founders is an ex-manager at Megafon; former members of the operator’s RnD subdivision have also joined the company. It is an independent technological partner of Megafon.