The comparison of autoencoder architectures in improving of prediction models

Abstract

In our day many prediction models require to encode the series of events in a way that will allow to train the model and obtain the highest quality of predictions. The encoding of events depends on data domain and applied methods, however one can use neural network to encode the series of actions and obtain informative features for predictive models. We compared several architectures of neural networks in a task of feature extraction for predictive models. The comparison of architectures of neural networks was obtained on the field of sequence modeling, where the popularity of LSTM networks is dominant. We found, that the usage of appropriate event encoding allows to improve the quality of CNN based networks without using the modification of architectures.