Naslov
An Exploration of CNN-based Time-series Prediction

Autori
Hrga, Ingrid

Vrsta, podvrsta i kategorija rada
Sažeci sa skupova, sažetak, znanstveni

Izvornik
Book of Abstracts 17th International Conference on Operational Research KOI 2018 / - , 2018, 148-148

Skup
17th International Conference on Operational Research (KOI 2018)

Mjesto i datum
Zadar, Hrvatska, 26.09.2018. - 28.09.2018

Vrsta sudjelovanja
Predavanje

Vrsta recenzije
Međunarodna recenzija

Ključne riječi
time-series prediction, convolutional neural networks, deep learning

Sažetak
Time-series prediction is considered for a long time an attractive research topic since the data exhibiting temporal dynamics is generated wherever there is a sign of human activity, with important applications ranging from healthcare, finance, economy or industry, to name just a few. However, due to their often high-nonlinearity and low signal-to-noise ratio, modeling and prediction of real-world data remains challenging. Moreover, large volumes of continuously generated data, along with the business requirements for their rapid analysis, call for a fully automated process which can be greatly enhanced by modern representation leaning methods based on deep neural networks. Over the last few years, convolutional neural networks (CNNs) have revolutionized the field of computer vision. Favored by the availability of the GPU hardware, which allowed parallelization of computations, and large labeled datasets, CNNs ability to automatically learn hierarchical feature representations directly from raw image data has led to results even surpassing human performance on some tasks. This has encouraged researchers to adopt them also on tasks of sequential modeling, such as machine translation or music composition, where recurrent neural networks (RNNs) were usually considered as a natural choice, and there is an increasing interest in building end-to- end time-series prediction systems with CNNs as their main building blocks. In this paper, we explore CNN-based approaches for time-series prediction. In line with the recent research, we follow two paths: in the first, time series are encoded as images, which enables direct application of image classification techniques. The second approach adapts the CNN architecture to specificities of time- series processing. The performance is tested on real-world data from different domains as choosing the appropriate approach for a specific task can contribute to better decision-making.

Izvorni jezik
Engleski

Znanstvena područja
Računarstvo, Informacijske i komunikacijske znanosti

POVEZANOST RADA