13 Aralık 2019, Cuma

Autoencoders

Autoencoder neural network is a popular ANN architecture applied mainly in unsupervised learning applications such anomaly detection. The most two applications that Autoencoders are applied for are fraud detection in bank transactions and dimensionality reduction like Principal Component Analysis. In other words, the core function of autoencoders is to teach the machines what constitutes “normal” data.

There are many other architectures of deep neural networks but these three are the most popular and widely-used ones besides to typical neural networks which have one input layer, multiple hidden layers and output layer. Our porposal as it is explained in details in section 4 follows the typical architecture of deep neuarl networks. The reason of chosing typical architecture is that our problem is regression problem so CNN is not be suitable for our case. CNN could be used in our case for extracting features from the dataset but we have applied features engineering by analyzing the most important technical indicators that have big impacts over the stock prices and our chosen features have been proved by the most well- known investing web portal www.investing.com and by the most popular economic traders. RNN is not suitable for our problem either because feedback connections in hidden layers leads to overfit in stock market prediction as the purpose is not remembering from old data but adjusting weights only. Finally we did not use Autoencoders as their main application is anomaly detecting and unsupervised learning but our case is supervised learning of regression type where in the training set we have label class representing the target stock price for the corresponding trading day.

There are many other architectures of deep neural networks but these three are the most popular and widely-used ones besides to typical neural networks which have one input layer, multiple hidden layers and output layer.

Autoencoders architecture.