- Title
- A framework to measure human behaviour whilst reading
- Creator
- Salehzadeh, Seyed Amirsaleh
- Creator
- Greyling, Jean
- Subject
- Computational intelligence
- Subject
- Machine learning Artificial intelligence Neural networks (Computer science)
- Date Issued
- 2019
- Date
- 2019
- Type
- Thesis
- Type
- Doctoral
- Type
- DPhil
- Identifier
- http://hdl.handle.net/10948/43578
- Identifier
- vital:36921
- Description
- The brain is the most complex object in the known universe that gives a sense of being to humans and characterises human behaviour. Building models of brain functions is perhaps the most fascinating scientific challenge in the 21st century. Reading is a significant cognitive process in the human brain that plays a critical role in the vital process of learning and in performing some daily activities. The study of human behaviour during reading has been an area of interest for researchers in different fields of science. This thesis is based upon providing a novel framework, called ARSAT (Assisting Researchers in the Selection of Appropriate Technologies), that measures the behaviour of humans when reading text. The ARSAT framework aims at assisting researchers in the selection and application of appropriate technologies to measure the behaviour of a person who is reading text. The ARSAT framework will assist to researchers who investigate the reading process and find it difficult to select appropriate theories, metrics, data collection methods and data analytics techniques. The ARSAT framework enhances the ability of its users to select appropriate metrics indicating the effective factors on the characterisation of different aspects of human behaviour during the reading process. As will be shown in this research study, human behaviour is characterised by a complicated interplay of action, cognition and emotion. The ARSAT framework also facilitates selecting appropriate sensory technologies that can be used to monitor and collect data for the metrics. Moreover, this research study will introduce BehaveNet, a novel Deep Learning modelling approach, which can be used for training Deep Learning models of human behaviour from the sensory data collected. In this thesis, a comprehensive literature study is presented that was conducted to acquire adequate knowledge for designing the ARSAT framework. In order to identify the contributing factors that affect the reading process, an overview of some existing theories of the reading process is provided. Furthermore, a number of sensory technologies and techniques that can be applied to monitoring the changes in the metrics indicating the factors are also demonstrated. Only, the technologies that are commercially available on the market are recommended by the ARSAT framework. A variety of Machine Learning techniques were also investigated when designing the BehaveNet. The BehaveNet takes advantage of the complementarity of Convolutional Neural Networks, Long Short-Term Memory networks and Deep Neural Networks. The design of a Human Behaviour Monitoring System (HBMS), by utilising the ARSAT framework for recognising three attention-seeking activities of humans, is also presented in this research study. Reading printed text, as well as speaking out loudly and watching a programme on TV were proposed as activities that a person unintentionally may shift his/her attention from reading into distractions. Between sensory devices recommended by the ARSAT framework, the Muse headband which is an Electroencephalography (EEG) and head motion-sensing wearable device, was selected to track the forehead EEG and a person’s head movements. The EEG and 3-axes accelerometer data were recorded from eight participants when they read printed text, as well as the time they performed two other activities. An imbalanced dataset consisting over 1.2 million rows of noisy data was created and used to build a model of the activities (60% training and 20% validating data) and evaluating the model (20% of the data). The efficiency of the framework is demonstrated by comparing the performance of the models built by utilising the BehaveNet, with the models built by utilising a number of competing Deep Learning models for raw EEG and accelerometer data, that have attained state-of-the-art performance. The classification results are evaluated by some metrics including the classification accuracy, F1 score, confusion matrix, Receiver Operating Characteristic curve, and Area under Curve (AUC) score. By considering the results, the BehaveNet contributed to the body of knowledge as an approach for measuring human behaviour by using sensory devices. In comparison with the performance of the other models, the models built by utilising the BehaveNet, attained better performance when classifying data of two EEG channels (Accuracy = 95%; AUC=0.99; F1 = 0.95), data of a single EEG channel (Accuracy = 85%; AUC=0.96; F1 = 0.83), accelerometer data (Accuracy = 81%; AUC = 0.9; F1 = 0.76) and all of the data in the dataset (Accuracy = 97%; AUC = 0.99; F1 = 0.96). The dataset and the source code of this project are also published on the Internet to help the science community. The Muse headband is also shown to be an economical and standard wearable device that can be successfully used in behavioural research.
- Format
- xv, 248 leaves
- Format
- Publisher
- Nelson Mandela University
- Publisher
- Faculty of Science
- Language
- English
- Rights
- Nelson Mandela University
- Hits: 1779
- Visitors: 1727
- Downloads: 74
Collections
NMU School of Computer Science, Mathematics, Physics and Statistics
| Thumbnail | File | Description | Size | Format | |||
|---|---|---|---|---|---|---|---|
| View Details Download | SOURCE1 | A Framework to Measure Human Behaviour whilst Reading.pdf | 17 MB | Adobe Acrobat PDF | View Details Download |