Publications

Detailed Information

Time Series Data Analysis using Deep Learning in Industry : 산업에서 딥러닝을 이용한 시계열 데이터 분석

DC Field Value Language
dc.contributor.advisor강명주-
dc.contributor.author김상연-
dc.date.accessioned2019-05-07T07:02:18Z-
dc.date.available2019-05-07T07:02:18Z-
dc.date.issued2019-02-
dc.identifier.other000000154575-
dc.identifier.urihttps://hdl.handle.net/10371/152943-
dc.description학위논문 (박사)-- 서울대학교 대학원 : 자연과학대학 협동과정 계산과학전공, 2019. 2. 강명주.-
dc.description.abstract딥러닝은 최근 몇 년간 산업 수학에 있어서 가장 강력하고 중요시 여겨지는 방법이다. 우리는 산업 수학의 시계열 데이터에 있어서 딥러닝 모델을 분석 및 예측 등에 정의하였다. 첫 번째로, 이상 감지를 위한 새로운 딥러닝 모델을 개발하였으며 이는 다양한 길이 뿐만 아니라 노이즈, 시간 차가 있는 데이터에서도 엔지니어에게 필요한 시계열 데이터 분석을 할 수 있었다. 두 번째로, 금융 시장의 트렌드를 예측하기 위한 다양한 딥러닝 모델을 개발 및 시험해보았으며 이 중 가중치 어텐션 네트워크의 경우 높은 예측 정확도 뿐만 아니라 시각화를 통해 직관적으로 모델을 이해 및 예측한 이유를 분석할 수 있었다.-
dc.description.abstractDeep learning, also called as artificial neural networks, is one of the most important and powerful subjects in industrial in recent years. Deep learning starts to show a great performance from image classification and in these days it have been applied to fields including computer vision, natural language process, speech recognition and etc. The performance is better than not only previous machine learning techniques, but also human experts in some cases. For an area with time series data, recurrent neural networks is widely used algorithm of deep learning. The aim of this theseis is to apply deep learning, especially with recurrent neural networks, for an industrial such as anomaly detection and trend prediction in financial market, with time series data . Its main contributions are (1) a new model for anomaly detection in time series data even for various length inputs, (2) various neural architectures for prediction in finance, and (3) attention networks and model analysis with attention vectors. Each experimental results of applications show better performances than previous machine learning techniques.-
dc.description.tableofcontents1 Introduction

2 Deep Learning Background

2.1 Neural Networks

2.2 Various Activation Functions

2.3 Error Backpropagation

2.4 Regularization

2.4.1 Dropout

2.4.2 Batch Normalization

3 Deep Learning Models

3.1 Multi Layer Perceptron

3.2 Convolutional Neural Networks

3.3 Recurrent Neural Networks

3.4 Long Short Term Memory

3.5 Attention Networks

4 Anomaly Detection

4.1 Related Works of Anomaly Detection

4.1.1 Anomaly detection

4.1.2 t-SNE

4.1.3 Clustering

4.2 Deep Correlation Mapping

4.2.1 LSTM

4.2.2 t-SNE

4.2.3 Full Model Architecture

4.2.4 Anomaly detection using Deep Correlation Mapping

4.3 Experimental Results

4.3.1 Correlation

4.3.2 Anomaly detection using DeepCorr

4.4 Conclusion

5 Trend Prediction

5.1 Related works of Trend Prediction

5.2 Trend Prediction with Deep Learning Models

5.2.1 Dataset

5.2.2 MLP

5.2.3 1D-CNN

5.2.4 LSTM

5.2.5 Attention Networks

5.2.6 Weighted Attention Networks

5.3 Experimental Results

5.3.1 Best Lookback Days

5.3.2 Results of Various Deep Learning Models

5.3.3 Visualization Attention Vectors

5.4 Conclusion

6 Conclusion and Future Works

Abstract (in Korean)

Acknowledgement (in Korean)
-
dc.language.isoeng-
dc.publisher서울대학교 대학원-
dc.subject.ddc004-
dc.titleTime Series Data Analysis using Deep Learning in Industry-
dc.title.alternative산업에서 딥러닝을 이용한 시계열 데이터 분석-
dc.typeThesis-
dc.typeDissertation-
dc.contributor.AlternativeAuthorSangyeon Kim-
dc.description.degreeDoctor-
dc.contributor.affiliation자연과학대학 협동과정 계산과학전공-
dc.date.awarded2019-02-
dc.contributor.majorDeep Learning-
dc.identifier.uciI804:11032-000000154575-
dc.identifier.holdings000000000026▲000000000039▲000000154575▲-
Appears in Collections:
Files in This Item:

Altmetrics

Item View & Download Count

  • mendeley

Items in S-Space are protected by copyright, with all rights reserved, unless otherwise indicated.

Share