Use this URL to cite or link to this record in EThOS:
Title: Interference reduction in classification and forecasting tasks through cluster and trend analysis
Author: Afolabi, David Olalekan
ISNI:       0000 0004 7658 8445
Awarding Body: University of Liverpool
Current Institution: University of Liverpool
Date of Award: 2018
Availability of Full Text:
Access from EThOS:
Access from Institution:
In order to classify or predict accurately in classification or time series tasks, the model building process is substantially dependent on the quality of data which is available for training such models. Consequently, reduced performance can be observed when input attributes/patterns have a conflicting influence on the learning process either due to intrinsic weak discrimination factor of some specific input attributes or as a result of outliers/anomaly picked up during data acquisition/entry. Several hypotheses are proposed, defined, and empirically tested to achieve an interference-less machine learning process using meta-assisted learning in data classification and time series forecasting. Meta-learning is a branch of machine learning that focuses on the automatic and flexible learning of informative concepts/knowledge mined from given data in an efficient manner to improve performance whereby such a system includes a process to monitor the learning progress. The two domains in which this research is focused on are classification tasks and time series forecasting tasks. Within these two domains, two further learning methods are explored whereby both the traditionally flat artificial neural network models and hierarchical structured artificial neural network models are modified to tackle the machine learning interference problem by using derived meta-information to reduce classification and forecasting error. The simulation experiments are performed with the multi-layer perceptron and a variant known as the constructive backpropagation artificial neural network for classification tasks; similarly, the nonlinear autoregressive exogenous model and long short-term memory artificial neural networks are used in time series forecasting tasks. This thesis is established on the following key hypotheses: i. Utilising the 'cluster assumption' for noise identification and extraction based on the intuition that samples of the dataset with higher similarity are inextricable and therefore should be clustered with other neighbouring samples that have similar labels. Clustered data from algorithms such as density based spatial clustering application with noise are analysed and are essential for the derivation of metainformation. ii. Detection of repeating trend patterns by decomposing input signal into several building-block components over a range of frequencies enables distinction between information and error/noise/anomaly. To filter or decompose time series trends, we apply the moving average and empirical mode decomposition respectively. iii. The guided meta-learning process; in which techniques are derived and introduced into the traditional learning process based on the inherent structure/distribution of pattern clusters or component signal trends within the data to tackle the problem of interference and noise within input attributes as the modified machine learner builds an accurate model. iv. Hierarchical learning of local and global clusters/trends as real-world information tends to be structured in a hierarchy of concepts. Therefore, it is intuitive to learn on small/uncomplicated clusters before tackling a complex/encompassing cluster; or in the case of time series, learning short-term patterns before long-term trends. This novel approach to noise elimination is shown to statistically increase the performance of a machine learning algorithm which is modified to carry out metalearning on the training data. It is applicable to various benchmark and real-world datasets with significant improvement on data that contains known/unknown structure or patterns. Therefore the methods put forward in this thesis have the potential to complement or reinforce existing machining learning algorithms.
Supervisor: Guan, Steven Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral