Using neural networks in the problem of classifying anomalous behavior in financial transactions using Python and Keras

Abstract

The purpose of this work is to study and develop neural networks for the problems of classifying anomalous behavior in financial transactions using Python and Keras. To achieve this goal, a dataset was chosen, consisting of 100,000 financial transactions, data was preprocessed, the architecture of the neural network was designed, and the process of its training was disclosed. The basis of the developed architecture is the input layer, which consists of 30 neurons corresponding to 30 features in our data. The output layer of the architecture is two layers that confirm or deny fraudulent transactions. The basis of the output layer is the softmax activation function, which generates probabilities for each class. One of the most important advantages of the developed neural network is the efficiency of activation functions, which is based on the use of ReLU (Rectified Linear Unit) for the input and hidden layers. The model was trained for 50 epochs using the Adam optimization algorithm and a batch size of 32. Adam was chosen as the optimizer, which provides faster and more stable convergence. Based on the results of training in 50 epochs, an accuracy of 88% was achieved. It is important to emphasize that, despite the significant imbalance of classes in the dataset, the model showed the ability to accurately detect anomalies. This opens up great opportunities for applying this approach in a real environment, where fraudulent transactions also usually make up a small proportion of the total number of transactions.