School of Computing and Information Sciences
Victor Potapenko is a Ph.D. candidate in the School of Computing and Information Sciences at Florida International University. He has an MBA from University of Miami with concentration in Finance, and an M.S. in Computer Science degree from Nova Southeastern University with concentration in Software Engineering. Victor is carrying out research on commercial applications of deep learning under the supervision of Dr. Naphtali Rishe. His current research interest is in predictive analytics of consumer behavior with deep learning models.
The complexity of modeling entity behavior with deep learning is proportional to sophistication of entity decision making ability within a given environment. In a controlled setting, environment and entity profile features are finite and known a priori. In the real world, entities’ behavior and environments exhibit a wide range of complexities represented by few, or virtually unlimited collections of features. Multiplicative complexity of dynamic environments and entities that adjust their behavior based on intrinsic and extrinsic factors presents a unique challenge for deep learning models. As the number of features used to represent this complexity grows, storage and computational requirements explode, and neural networks struggle to learn meaningful representations from data during training.
Customer attrition prediction is a problem of predicting a facet of behavior of a complex entity in a complex dynamic environment. Customer’s actions are a function of past behavior as well as macroeconomic and microeconomic environments in time space. A wide spectrum of deep learning architectures can be used to learn this function. The proposed deep learning approach is targeted to solve this complexity by learning compact embeddings of dynamic environments from large sets of features and augmenting model with embeddings during training and inference. We contend that these embeddings representing relative features of customer environment and profile improve inferential power of deep learning models that predict customer attrition and reduce model complexity.