Computing and Information Systems - Theses
Permanent URI for this collection
Search Results
1 - 10 of 190
-
ItemScalable and Explainable Time Series Classification( 2022)Time series data are ubiquitous, and the explosion of sensor technologies on industry applications has further accelerated their growth. Modern applications collecting large datasets with long time series require fast automated decisions. Besides, legislation such as the European General Data Protection Regulation now require explanations for automated decisions. Although state-of-the-art time series classification methods are highly accurate, their computationally expensive and complex models are typically impractical for large datasets or cannot provide explanations. To address these issues, this thesis proposes two time series classification methods that are comparable to state-of-the-art methods in terms of accuracy, while the two methods provide scalable and explainable classifications. Our first method proposes a novel supervised selection of sub-series to pre-compute a set of features that maximizes the classification accuracy when fed into a tree-based ensemble. Our second method further proposes a perturbation scheme for the supervised feature selection and the node-splitting process when training the tree-based ensemble. We also propose a highly time-efficient strategy to build the tree-based ensemble. Both methods enable explainability for our classification results, while they are orders of magnitude faster than the state-of-the-art methods. Our second method, in particular, is significantly faster and more accurate than our first approach, while it is not significantly different from the state-of-the-art methods in terms of classification accuracy. Moreover, motivated to explore a more general model for time series classification, we propose a novel graph-based method to learn to classify time series without the order constraint inherent to time series data. This method classifies time series by learning the relationships among the data points independently from their positions (i.e., time-stamps) within the series. We show that this method outperforms state-of-the-art methods over several time series datasets, thus opening up a new direction for the design of time series classifiers.
-
ItemEnhancing Mid-Air Selection and Manipulation for Complex Virtual Reality Interaction( 2023)Object selection and manipulation are fundamental to interacting with objects in Virtual Reality (VR) systems. Existing object selection and manipulation techniques in VR are primarily based on mid-air interaction with virtual hands or ray pointers. They are simple and intuitive but are often criticized in the literature for being imprecise, inefficient, and cumbersome. Specifically, these techniques are insufficient for complex VR interaction scenarios that contain small, distant, and occluded targets and require efficient, precise, versatile, and prolonged operations. This thesis presents occlusion visualization techniques, integration strategies of complementary modalities of gaze and body surfaces, and predictive systems based on target prediction models to enhance virtual hands and ray pointers for complex VR interactions. Findings from a series of user studies demonstrated that the proposed solutions could select and manipulate small, distant, and occluded targets in an effective, efficient, comfortable, and satisfying manner. Overall, our technical solutions and findings can inform the future design of more usable and useful 3D user interfaces for VR systems.
-
ItemMethodological Issues in the Development of Cognition-Aware Attention Management Systems in HCI( 2023)Attention is one of humans’ most critical resources and the ability to manage it directly impacts our productivity level. Attention management systems have been proposed to keep track of individuals’ attentional states in an attempt to moderate and design effective interventions. The first step in designing such systems is to detect users’ attentional state automatically and unobtrusively. Recently, the potential of physiological signals to measure humans’ internal states has drawn a great amount of attention toward them in Human-Computer Interaction (HCI); however, the intricacy in the collection and analysis of these signals and ambiguity in the concept of attention threaten the validity of studies in this area. Therefore, in this thesis, we aimed to shed light on methodological issues that exist in the development of attention management systems using physiological signals. To achieve this goal, first, in an observational study, we attempted to use these signals to measure attentional state with the aim to spot improprieties of HCI methods in this area. This study that is presented in Chapter 3 highlighted the existence of severe methodological issues in the collection and analysis of physiological signals and the theoretical underpinning of attentional state constructs and their subjective scales. Then, to refine the methodological issues that exist in HCI literature regarding the use of physiological signals, we developed a systematic critical review of CHI studies using Electrodermal Activity (EDA), one of the physiological signals that have been used frequently in HCI to detect internal states. This review which is presented in Chapter 4 highlighted severe methodological issues in CHI EDA practices such as a lack of transparency, the incorrect use of recording equipment, poor controls in study designs, and the misuse of signal processing techniques and analysis procedures. To refine these issues we developed a set of guidelines that ensures the minimum required quality of an EDA practice. Furthermore, to raise HCI’s understanding of the concept of attentional state, we conducted two reviews and a validation experiment on Mental Workload (MWL), one of the widely used attentional state-related constructs in HCI. In these studies which are presented in Chapter 5, we investigated MWL’s definitions, theories, applications at CHI, and its subjective scales. Our findings stressed the obscurity of this construct and severe flaws in its subjective scales and their unfitness for HCI tasks. This thesis facilitates the development of attention management systems using physiological signals by contributing to better science and the broader and more systematic use of physiological signals and attentional state constructs and emphasizes the urgent need for an increased awareness of methodological standards in HCI.
-
ItemTraffic Optimization with Deep Reinforcement Learning in a Connected Transport Ecosystem( 2022)The emergence of connected autonomous vehicles and smart traffic infrastructure provides an opportunity to manage traffic using novel techniques as the data generated from these entities can be used for intelligent real-time traffic control. Reinforcement learning has been widely adopted in various real-time control and sequential decision-making tasks including robotics, health care, NLP, and games due to the ability to control complex environments using observed data. Despite the success of reinforcement learning, applying off-the-shelf reinforcement algorithms for novel traffic management solutions is difficult due to the scalability, dynamicity, graph-structured data and multi-objectivity nature present in these transportation problems. In this thesis, we propose several solutions for intelligent traffic management control using novel reinforcement learning algorithms and architectures to overcome the above limitations. The thesis starts by exploring novel reinforcement learning-based solutions for intersection level traffic management problems and proposes a multi-agent solution with a novel reinforcement learning algorithm to handle multi-objectivity. Then, the thesis expands to road network-level traffic management problems and proposes a coordination-based reinforcement learning architecture to tackle the existing limitations in the road network level traffic optimization. Finally, we propose a generalized reinforcement learning architecture for common combinatorial graph problems. This enables many intelligent problems in the transportation sector to be solved using a single reinforcement learning framework. This generalized reinforcement learning framework is able to outperform the state-of-the-art on a variety of dynamic graph problems and can be applied to several traffic management solutions in a connected transport system. Based on the research outcomes, it is evident that using novel deep reinforcement learning algorithms for novel traffic management techniques has several advantages, making deep reinforcement learning as the go-to solution for the connected transport ecosystem.
-
ItemEnhancing Deep Multimodal Representation: Online, Noise-robust and Unsupervised Learning( 2022)Information that is generated and shared today uses data that involves different modalities. These multimodalities are not limited to the well-known sensory media (e.g., text, image, video, and audio), but could be any abstract or inferred form of encoding information (e.g., propagation network of a news article and sentiment of a text) that represents a different viewpoint of the same object. For machine learning models to be competitive with humans, they should be able to extract and combine information from these modalities. Thus, multimodal representation learning has emerged as a broad research domain that aims to understand complex multimodal environments while narrowing the heterogeneity gap among different modalities. Due to the potential of representing latent information in complex data structures, deep learning-based techniques have recently attracted much attention for multimodal representation learning. Nevertheless, most existing deep multimodal representation learning techniques lack the following: (1) ability to continuously learn and update representations in a memory-efficient manner while being recency-aware and avoiding catastrophic forgetting of historical knowledge; (2) ability to learn unsupervised representations for under-exploited multimodalities with complex data structures (i.e., temporally evolving networks) and high diversity (cross-domain multimodal data); and (3) ability to directly serve as features to address various real-world applications without fine-tuning using an application-specific labelled dataset. This thesis aims to bridge these research gaps in deep multimodal representation learning approaches. In addition, this thesis addresses real-world applications involving multimodal data such as misinformation detection, spatiotemporal activity modeling and online market basket analysis. The main contributions of this thesis include: (1) proposing two novel online learning strategies for learning deep multimodal representations, and proposing two frameworks using the proposed online learning strategies to address two real-world applications -- i.e., user-guided spatiotemporal activity modeling (USTAR) and online market basket analysis (OMBA); (2) proposing METEOR, a memory and time efficient online representation learning algorithm for making deep multimodal representations compact and scalable to cope with the different data rates of real-world multimodal data streams; (3) developing an unsupervised framework to capture and preserve domain-specific and domain-shared knowledge in cross-domain data streams, and applying the proposed framework to address cross-domain fake news detection; (4) proposing an unsupervised model to learn representations for temporally evolving graphs by mimicking the future knowledge of an evolving graph at an early timestep, and developing a new framework called Propagation2Vec with the help of the proposed objective functions for fake news early detection; and (5) developing a theoretically-motivated noise-robust unsupervised learning framework, which can filter out the noise (i.e., fine-tune) in multimodal representations learned from general pretraining objective functions without requiring a labelled dataset, and applying the findings to address the unsupervised fake news detection task.
-
ItemDetection of Repeating Activities Recorded by Sensors in the Absence of Labeled Data( 2022)Time series data with repetition is generated in many applications such as remote patient monitoring using wearable sensors. The consecutive repeating patterns in a time series indicate that an event is re-occurring consecutively over time. As an example, a patient’s rehabilitation exercise movements collected by an accelerometer can determine whether the patient is repeatedly exercising. In such applications collecting and labelling real data is difficult, lengthy and expensive. The exercises also might be performed differently by different patients and even one patient might start to perform the exercise in a certain way but the movement can change due to for example fatigue, thus generating varying repeating patterns in the underlying time series. In this thesis, we have explored pattern mining, machine learning and deep learning techniques that can be used for detecting consecutive repeating patterns from a time series without previous knowledge about the repeats and without use of labelled data. We have shown the limitations of the defacto benchmark called the Matrix Profile, for detecting consecutive repeating patterns and have addressed those limitations by modifying Matrix Profile and proving how to set its key input parameter. We have created a transferable deep learning technique, RP-Mask, for detecting and localising segments of repeating patterns on a time series, which learns entirely from synthetic data and is able to transfer this learning to an unseen real dataset. We further explore whether RP-Mask is using the correct information in the time series to make decisions. We show that the state-of-the-art explainability techniques generate different explanations for the same model and create a new explainability approach using a low pass frequency filter on the time series.
-
ItemBandit Algorithms with Application to Database Index Selection( 2023)Indices are data structures used by databases to support efficient data retrieval during query executions. However, due to computational and space limits, only a subset of all possible indices are created in practice. The selection of such a set is currently managed by human database administrators (DBAs) with the help of offline tools provided by the database vendors. Unfortunately, with today’s complex workloads, offline tools fail to provide accurate cost estimates or optimal index recommendations, contributing to poor or costly configurations by DBAs. In this thesis, we aim to devise a machine learning-based recommendation tool for a performant subset of indices formulated as a sequential decision-making problem under uncertainty, specifically under the multi-armed bandit (MAB) setting. In this setting, a set of possible actions (or arms) are presented, one of which must be played each round by a bandit learner. A stochastic reward is observed by the bandit, while no feedback is provided for the unplayed actions. To maximise cumulative rewards, it is imperative for a bandit learner to maintain a balance between exploration and exploitation. Owing to the setting’s simplicity, it is possible to prove cumulative reward bounds for many variants of the basic MAB setting. One such MAB variant is named the Contextual Combinatorial Multi-Armed Bandit, which provides a pathway for us to choose multiple actions in a round where each action’s side information is available. This can be adapted to our problem via a careful design of context feature vectors, producing up to 75% faster response time as observed in dynamic database workloads as compared to a commercial tuning tool. In order to prevent poor performance in early rounds due to the need for the exploration, a way of warm-starting contextual bandits is also explored. This approach allows users to accommodate prior knowledge gained from past query executions. This knowledge is robustly combined to allow some mismatch between the past knowledge and the actual rounds. A set of data can be inexpensively obtained from existing offline tools, making it a natural choice for the pre-training dataset in the automatic index tuning application when past observations are unavailable. Unfortunately, these data come with artificial reward units not easily convertible to a more conventional one. This more significant mismatch makes these data sub-optimal as sources of existing knowledge. We propose a general framework for pre-training datasets with a latent scaling of reward units. A range of experiments demonstrates the effectiveness of our approach.
-
ItemDynamic Multi-objective Optimisation Using Evolutionary Algorithms( 2022)Dynamic Multi-objective Optimization Problems (DMOPs) offer an opportunity to examine and solve challenging real world scenarios where trade-off solutions between conflicting objectives change over time. Definition of benchmark problems allows modelling of industry scenarios across transport, power and communications networks, manufacturing and logistics. Recently, significant progress has been made in the variety and complexity of DMOP benchmarks and the incorporation of realistic dynamic characteristics. However, significant gaps still exist in standardised methodology for DMOPs, specific problem domain examples and in the understanding of the impacts and explanations of dynamic characteristics. This thesis provides major contributions on these three topics within evolutionary dynamic multi-objective optimization. Firstly, experimental protocols for DMOPs are varied. This limits the applicability and relevance of results produced and conclusions made in the field. A major source of the inconsistency lies in the parameters used to define specific problem instances being examined. The uninformed selection of these has historically held back understanding of their impacts and standardisation in experimental approach to these parameters in the multi-objective problem domain. Using the frequency and severity (or magnitude) of change events, a more informed approach to DMOP experimentation is conceptualized, implemented and evaluated. Establishment of a baseline performance expectation across a comprehensive range of dynamic instances for well-studied DMOP benchmarks is analyzed. To maximize relevance, these profiles are composed from the performance of evolutionary algorithms commonly used for baseline comparisons and those with simple dynamic responses. Comparison and contrast with the coverage of parameter combinations in the sampled literature highlights the importance of these contributions. Secondly, the provision of useful and realistic DMOPs in the combinatorial domain is limited in previous literature. A novel dynamic benchmark problem is presented by the extension of the Travelling Thief Problem (TTP) to include a variety of realistic and contextually justified dynamic changes. Investigation of problem information exploitation and it's potential application as a dynamic response is a key output of these results and context is provided through comparison to results obtained by adapting existing TTP heuristics. Observation driven iterative development prompted the investigation of multi-population island model strategies, together with improvements in the approaches to accurately describe and compare the performance of algorithm models for DMOPs, a contribution which is applicable beyond the dynamic TTP. Thirdly, the purpose of DMOPs is to reconstruct realistic scenarios, or features from them, to allow for experimentation and development of better optimization algorithms. However, numerous important characteristics from real systems still require implementation and will drive research and development of algorithms and mechanisms to handle these industrially relevant problem classes. The novel challenges associated with these implementations are significant and diverse, even for a simple development such as consideration of DMOPs with multiple time dependencies. Real world systems with dynamics are likely to contain multiple temporally changing aspects, particularly in energy and transport domains. Problems with more than one dynamic problem component allow for asynchronous changes and a differing severity between components that leads to an explosion in the size of the possible dynamic instance space. Both continuous and combinatorial problem domains require structured investigation into the best practices for experimental design, algorithm application and performance measurement, comparison and visualization. Highlighting the challenges, the key requirements for effective progress and recommendations on experimentation are explored here.
-
ItemImproving Agile Sprint Planning Through Empirical Studies of Documented Information and Story Points Estimation( 2022)In Agile iterative development (e.g., Scrum), effort estimation is an integral part of the development iteration planning (i.e., sprint planning). Unlike traditional software development teams, an Agile team relies on a lightweight estimation method based on the team consensus (e.g., Planning Poker) for effort estimation and the estimated effort is continuously refined (or changed) to improve the estimation accuracy. However, such lightweight estimation methods are prone to be inaccurate and late changes of the estimated effort may cause the sprint plan to become unreliable. Despite a large body of research, only few studies have reviewed the reasons for inaccurate estimations and the approaches to improve effort estimation. We conducted a systematic literature review and found that the quality of the available information is one of the most common reasons for inaccurate estimations. We found several manual approaches aim to help the team improve the information quality and manage the uncertainty in effort estimation. However, prior work reported that the practitioners were reluctant to use them as they added additional overhead to the development process. The goals of this thesis are to better understand and propose the approaches that help the team achieves accurate estimation without introducing additional overhead. To achieve this goal, we conducted studies in this thesis in two broad areas. We first conducted two empirical studies to investigate the importance of documented information for effort estimation and the impact of estimation changes in a project. In the first empirical study, we aim to investigate the importance and quality of documented information for effort estimation. We conducted a survey study with 121 Agile practitioners from 25 countries. We found that the documented information is considered important for effort estimation. We also found that the useful documented information for effort estimation is often changed and the practitioners would re-estimate effort when the change of documented information occurred, even after the work had started. In the second empirical study, we aim to better understand the change of effort (in Story Points unit; SP). We examined the prevalence of SP changes, the accuracy of changed SP, and the impact of information changes on SP changes. We found that the SP were not often changed after sprint planning. However, when the SP were changed, the changing size was relatively large and the changed SP may be inaccurate. We also found that the SP changes were often occurred along with the information changes for scope modification. These findings suggest that a change of documented information could lead to a change of effort, and the changed effort could have a large impact on the sprint plan. To mitigate the risk of an unreliable sprint plan, the documented information and the estimated effort should be verified and stabilized before finalizing the sprint plan. Otherwise, the team may have to re-estimate the effort and adjust the sprint plan. However, revisiting all documented information and estimated SP could be a labor-intensive task and may not comply with the Agile principles. To help the team manages these uncertainties without introducing additional overhead, we proposed the automated approaches called DocWarn and SPWarn to predict the documentation changes and SP changes that may occur after sprint planning. We built DocWarn and SPWarn using machine learning and deep learning techniques based on the metrics that measure the characteristics of the work items. We evaluated DocWarn and SPWarn using the work items extracted from the open-source projects. Our empirical evaluations show that DocWarn achieved an average AUC of 0.75 and SPWarn achieved an average AUC of 0.73, which are significantly higher than baseline models. These results suggest that our approaches can predict future changes of documented information and SP based on the currently-available information. With our approaches, the team will be better aware and pay attention to the potential documentation changes and SP changes during sprint planning. Thus, the team can manage uncertainty and reduce the risk of unreliable effort estimation and sprint planning without additional overhead.
-
ItemA Novel Perspective on Robustness in Deep Learning( 2022)Nowadays, machine learning plays a crucial role in our path toward automated decision-making. Traditional machine learning algorithms would require careful, often manual, feature engineering to deliver satisfactory results. Deep Neural Networks (DNNs) have shown great promise in an attempt to automate this process. Today, DNNs are the primary candidate for various applications, from object detection to high-dimensional density estimation and beyond. Despite their impressive performance, DNNs are vulnerable to different security threats. For instance, in adversarial attacks, an adversary can alter the output of a DNN for its benefit by adding carefully crafted yet imperceptible distortions to clean samples. As another example, in backdoor (Trojan) attacks, an adversary intentionally plants a loophole in the DNN during the learning process. This is often done via attaching specific triggers to the benign samples during training such that the model creates an association between the trigger and a particularly intended output. Once such a loophole is planted, the attacker can activate the backdoor with the learned triggers and bypass the model. All these examples demonstrate the fragility of DNNs in their decision-making, which questions their widespread use in safety-critical applications such as autonomous driving. This thesis studies these vulnerabilities in DNNs from novel perspectives. To this end, we identify two key challenges in the previous studies around the robustness of neural networks. First, while a plethora of existing algorithms can robustify DNNs against attackers to some extent, these methods often lack the efficiency required for their use in real-world applications. Second, the true nature of these adversaries has been less studied, leading to unrealistic assumptions about their behavior. This is particularly crucial as building defense mechanisms using such assumptions would fail to address the underlying threats and create a false belief in the security of DNNs. This thesis studies the first challenge in the context of robust DNN training. In particular, we leverage the theory of coreset selection to form informative weighted subsets of data. We use this framework in two different settings. First, we develop an online algorithm for filtering poisonous data to prevent backdoor attacks. Specifically, we identify two critical properties of poisonous samples based on their gradient space and geometrical representation and define an appropriate selection objective based on these criteria to select clean samples. Second, we extend the idea of coreset selection to adversarial training of DNNs. Although adversarial training is one of the most effective methods in defending DNNs against adversarial attacks, it requires generating costly adversarial examples for each training sample iteratively. To ease the computational burden of various adversarial training methods in a unified manner, we build a weighted subset of the training data that can faithfully approximate the DNN gradient. We show how our proposed solution can lead to robust neural network training more efficiently in both of these scenarios. Then, we touch upon the second challenge and question the validity of one of the widely used assumptions around adversarial attacks. More precisely, it is often assumed that adversarial examples stem from an entirely different distribution than clean data. To challenge this assumption, we resort to generative modeling, particularly Normalizing Flows (NF). Using an NF model pre-trained on clean data, we demonstrate how one can create adversarial examples closely following the clean data distribution. We then use our approach against state-of-the-art adversarial example detection methods to show that methods that explicitly assume a difference in the distribution of adversarial attacks vs. clean data might greatly suffer. Our study reveals the importance of correct assumptions in treating adversarial threats. Finally, we extend the distribution modeling component of our adversarial attacker to increase its density estimation capabilities. In summary, this thesis advances the current state of robustness in deep learning by i) proposing more effective training algorithms against backdoor and adversarial attacks and ii) challenging a fundamental prevalent misconception about the distributional properties of adversarial threats. Through these contributions, we aim to help create more robust neural networks, which is crucial before their deployment in real-world applications. Our work is supported by theoretical analysis and experimental investigations based on publications.