|Title:||Grammar based feature generation for time-series prediction|
|Authors:||De Silva, Anthony Mihirana|
|Publisher:||University of Sydney.|
Graduate School of Engineering and IT
School of Electrical & Information Engineering
|Abstract:||The application of machine learning techniques to predict time-series continues to attract considerable attention due to the difficulty of the prediction problems compounded by the non-linear and non-stationary nature of the real world time-series. The performance of machine learning techniques, among other things, depends on suitable engineering of features. This thesis proposes a systematic way for generating suitable features using context-free grammar. The notion of grammar families as a compact representation to generate a broad class of features is exploited. Implementation issues and ways to overcome them are explained in detail. A number of feature selection criteria are investigated and a hybrid feature generation and selection algorithm using grammatical evolution is proposed. The proposed approaches are demonstrated by predicting the closing price of major stock market indices, peak electricity load and net hourly foreign exchange client trade volume. The widely and commonly employed features in practice (in previous work) for electricity and financial time-series are explored. These features are considered as a basis for comparison with the features generated and selected by the proposed framework. Other model-based approaches and naive approaches are also used as benchmarks. It is shown that the generated features can improve results, while requiring no domain-specific knowledge. The proposed method is used to determine suitable features to use in predicting previously unexplored foreign exchange client trade volume and the capabilities of the approach in automatically engineering appropriate features is highlighted. The proposed method can be applied to a wide range of machine learning architectures and applications to represent complex feature dependencies explicitly when machine learning cannot achieve this by itself.|
|Access Level:||Access is restricted to staff and students of the University of Sydney . UniKey credentials are required. Non university access may be obtained by visiting the University of Sydney Library.|
Access is restricted to staff and students of the University of Sydney . UniKey credentials are required. Non university access may be obtained by visiting the University of Sydney Library.
|Rights and Permissions:||The author retains copyright of this thesis. It may only be used for the purposes of research and study. It must not be used for any other purposes and may not be transmitted or shared with others without prior permission.|
|Type of Work:||Masters Thesis|
|Type of Publication:||Master of Philosophy M.Phil|
|Appears in Collections:||Sydney Digital Theses (University of Sydney Access only)|
|desilva_am_thesis.pdf||Thesis||1.52 MB||Adobe PDF|
Items in Sydney eScholarship Repository are protected by copyright, with all rights reserved, unless otherwise indicated.