Sergey is a Data Scientist at Zalando where he is working on both researching the possible improvements of machine learning models for fraud detection and on the infrastructure for training and serving the models. Previously he was a computer vision engineer in the research environment. Sergey's background consists of applied mathematics and physics along with computer science and data science.
The cold start problem is one of the biggest challenges when applying machine learning in industry. The goal is to build an accurate production-ready machine learning model when there is either no data or insufficient data available to train a model - the so-called cold start. In this case methods like transfer learning are proved to be valuable by using a model that was trained on a different but related task.In his talk Sergey will guide through a real-world cold start problem he was working on in payment fraud detection at Zalando. The business case aims at building a payment fraud detection model for a new market, where there is no historical data available. The challenge is how to combine the data from existing markets in an elaborate way so that it generalizes to a new market. Sergey will discuss several stages of rolling-out a model for the cold-start problem and possible solutions such as reliable cross-domain validation, transfer learning, domain adaptation and domain generalization.