![]() Since the outputs are real, the activation function at the output is linear. For the price forecasting problem, the target is multivariate – the outputs are the next day’s variables (open, close, high, low and NAV). Each of the inputs is processed by either TimeNet or ConvTimeNet and the features from the five channels are fused by a fully connected layer with an output target. ![]() They were converted to multivariate form owing to our requirement of feeding five types of inputs. These models were originally developed for univariate data. We have compared ConFuse with two state-of-the-art deep learning based time series models, namely TimeNet based on LSTM, and ConvTimeNet based on CNN. Comparisons with state-of-the-art methods illustrate the benefits of the proposed method. An original end-to-end training strategy is proposed, that takes advantage of a key property of activation functions, thus allowing the use of the efficient stochastic procedure for learning the architecture parameters. Each channel is hence processed by 1D CTL and the resulting representations are fused by a fully connected network of transform learning, leading to the so-called ConFuse approach. Motivated by the success of 1D CNNs for time series processing, coupled with the need for an unsupervised representation learning tool, we propose to make use of our recently introduced convolutional transform learning (CTL) approach. Instead of creating two separate end-to-end models, for performing classification and regression, we propose to learn a single unsupervised model. The former is a regression problem where the task is to predict the value of a stock, and the latter amounts to decide whether to buy or sell a stock. In this work, we focus on two important problems of stock analysis, namely stock forecasting and stock trading. In contrast, if the learning paradigm was unsupervised, only one network would be needed, providing as an output relevant features, which could, later on, serve as input to a third-party classifier or regressor. In the said work, two distinct deep neural networks are trained to solve both cases. Consider for instance the problem of ECG data classification, involving either four or sixteen classes. Such an unsupervised strategy may offer greater flexibility than supervised ones. A recent alternative has been proposed in, relying on an unsupervised method that learns feature representations from multi-channel data, and then uses these features as an input of a suitable classifier or regression method, adapted to the user’s need. It is worth mentioning that most existing works address inference from multi-channel data using a supervised end-to-end machine learning paradigm, where the input is the raw multi-channel signal and the output is the inferred parameter (e.g., class label, regression value). In several such studies, all the sensors are stacked one after the other to form a matrix and 2D CNN is used for analyzing these signals. Deep learning has also been widely used for analyzing multi-channel / multi-sensor signals.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |