Use of Neural Network Models for Learning Embeddings on Card Transaction Data of Bank Client for Different Tasks

Neural networks are actively used in various tasks. Neural networks are so popular because they can solve different problems: from modeling on flat tables to texts, images, music and video processing The Bank operates, collects and processes a large amount of customer operations data. Bank customers make a huge number of transactions. And transactional data has a temporary structure. In this research the hypothesis is verified that the sequence of transactions contains behavioral information about the client. The goal of this work is to extract meaningful information from transactional data on clients, in other words, training of hidden vector representations (embeddings) of client's transactional behavior. As part of this work, it is planned to train the LSTM autoencoder on bank card transactions data. The model will get sliding windows of consecutive transactions at the level of one client. The task of an autoencoder is to obtain available, accessible information about all transactions from incoming windows that are enough to recreate the incoming version with high accuracy. These embeddings and their derivatives will be used as variables in the presented bank models of estimating the probability of default of contracts on primary and secondary bank offers. These models have already contained simple aggregation variables that do not take into account time dependence. Using the LSTM autoencoder learnt embeddings allow to obtain sufficient uplifts to presented models (2.4-3.7 Gini). This indicates that this method allows you to extract time dependencies on transactions that can give the necessary client behavioral review.