Advances in artificial intelligence and machine learning have led to a steep rise in the adoption of AI to augment or support human decision-making across domains.
There has been an increasing body of work addressing the benefits of model interpretability and explanations to help end-users or other stakeholders decipher the inner workings of the so-called "black box AI systems".
Yet, little is currently understood about the role of modalities through which explanations can be communicated (e.g. text, visualizations, or audio) to inform, augment, and shape human decision-making.
This thesis addresses this research gap through the lens of a credibility assessment system.
Considering the deluge of information available through various channels, people constantly make decisions while considering the perceived credibility of the information they consume.
However, with an increasing information overload, assessing the credibility of the information we encounter is a non-trivial task.
To help users in this task, automated credibility assessment systems have been devised as decision support systems in various contexts (e.g. assessing the credibility of news or social media posts).
However, for these systems to be effective in supporting users, they need to be trusted and understood.
Explanations have been shown to play an essential role in informing users' reliance on decision support systems.
This thesis investigates the influence of explanation modalities on an AI-assisted credibility assessment task.
A between-subjects experiment (N=375) was performed, spanning six different explanation modalities, to evaluate the role of explanation modality on the accuracy of AI-assisted decision outcomes, the perceived system trust among users, and system usability.
The results indicate that explanations play a significant role in shaping users' reliance on the decision support system and, thereby, the accuracy of decisions made.
Users were found to perform with higher accuracy while assessing the credibility of statements in the presence of explanations.
Users additionally had a significantly harder time agreeing on statement credibility without explanations. With explanations present, text and audio explanations were more effective than graphic explanations.
This thesis concludes that combinations of graphical and text and/or audio explanations were significantly effective. Such combinations of modalities led to a higher user performance than using graphical explanations alone.

Explaining the behaviour of Artificial Intelligence models has become a necessity. Their opaqueness and fragility are not tolerable in high-stakes domains especially.
Although considerable progress is being made in the field of Explainable Artificial Intelligence, scholars have demonstrated limits and flaws of existing approaches: explanations requiring further interpretation, non-standardised explanatory format, and overall fragility.
In light of this fragmentation, we turn to the field of philosophy of science to understand what constitutes a good explanation, that is, a generalisation that covers both the actual outcome and, possibly multiple, counterfactual outcomes.
Inspired by this, we propose CHIME: a human-in-the-loop, post-hoc approach focused on creating such explanations by establishing the causal features in the input.
We first elicit people's cognitive abilities to understand what parts of the input the model might be attending to. Then, through Causal Discovery we uncover the underlying causal graph relating the different concepts. Finally, with such a causal structure, we compute the causal effects different concepts have on a model's outcome.
We evaluate the Fidelity, Coherence, and Accuracy of the explanations obtained with CHIME with respect to two state-of-the-art Computer Vision models trained on real-world image data sets. We found evidence that the explanations reflect the causal concepts tied to a model's prediction, both from the perspective of causal strength and accuracy.

While the performance of traditional confidence-based rejectors is heavily dependent on the calibration of the pretrained model, this study proposes the concept of feature-based rejectors and the whole pipeline where such rejector can be used in. Multiple design and development decisions along the implementation are discussed in the paper - involving humans in the loop of creating the feature set that machine should use and defining value, using interpretability methods such as saliency map to extract the features that the machine actually uses, and SceneRejector is the end product of the whole process. It is relevant to the field of computer vision, and it is applied to the task of scene classification within the scope of this paper. Its performance is evaluated against baselines with accuracy, rejection rate, and a new metric named value, which more comprehensively measures the practical value of machine learning with a reject option in different use cases. Experiments have proved the concept of a feature-based rejector works and it is able to filter out unknown unknowns, which is a challenge to the confidence-based rejectors. It also creates better value than the traditional confidence-based rejectors in some cases, especially when the machine learning model is not well calibrated. Further analysis is conducted to understand the behavior of SceneRejector as well. It is discovered that SceneRejector brings better value than confidence-based rejectors when the pretrained model is not well calibrated, and that rejection rate and accuracy of the rejector is also correlated with the value.

Crowd-powered conversational systems (CPCS) solicit the wisdom of crowds to quickly respond to on-demand users' needs. The very factors that make this a viable solution ---such as the availability of diverse crowd workers on-demand--- also lead to great challenges. The ever-changing pool of online workers powering conversations with individual users makes it particularly difficult to generate contextually consistent responses from a single user's standpoint. To tackle this, prior work has employed conversational facts extracted by workers to maintain a global memory, albeit with limited success. Leveraging systematic context in affective crowdsourcing tasks has remained unexplored. Through a controlled experiment, we explored if a conversational agent, dubbed ContextBot, can provide workers with the required context on the fly for successful completion of affective support tasks in CPCS, and explore the impact of ContextBot on the response quality of workers and their interaction experience. To this end, we recruited workers (N=351) from the Prolific crowdsourcing platform and carried out a 3*3 factorial between-subjects study. Experimental conditions varied based on (i) whether or not context was elicited and informed by motivational interviewing techniques (MI, non-MI, and chat history), and (ii) different conversational entry points for workers to produce responses (early, middle, and late). Our findings showed that workers who entered the conversation earliest were more likely to produce highly consistent responses after interacting with ContextBot. Better user experience from workers was expected after they interacted with ContextBot at a late entry. We found that interacting with ContextBot through task completion did not negatively impact workers' cognitive load.

Generating synthetic images has wide applications in several fields such as creating datasets for machine learning or using these images to investigate the behaviour of machine learning models. An essential requirement when generating images is to control aspects such as the entities or objects in the image. Controlling this helps in creating custom datasets tailored for the above applications and creates a platform to conduct diverse experiments with the generated images, enabling research in the application's field.Existing methods individually enable controllability over various elements in the image such as selecting the objects, their properties, colour or the relations between objects, etc. but we identify a research gap in this field where no single method allows the user to control all these aspects of the image. An additional research gap identified is that existing methods cannot generate images based on a query with a logic based combination of entities.

In this thesis, we aim to fill this research gap by developing SceneUI - a system that allows the user to specify and control aspects of the scene through a user interface such as the objects, object properties, spatial relations between objects, object colour, extent of contextual objects and the background of the image. Additionally, we include a component where the scene is generated based on an OR query specifying the objects as predicates, which serves as a foundation to generating images based on entity combinations. Owing to the limited range of attributes for objects and the lack of objects with attributes in the dataset, we augment the dataset by expanding the attributes of objects in the scene graphs of the dataset and introducing additional objects that have attributes. This was done by identifying recurring objects in the dataset that could be expanded and manually annotating the changes in the dataset. This increases the level of controllability and gives a wider range of an object's properties to choose from.

The goal of the thesis is to design and develop a method that allows the user to declare, specify and manipulate elements of the image and eventually use the images generated for two use cases - Generating Images for Interpretable Machine Learning and Generating Images from Queries as Ground Truth. We evaluate SceneUI to show its usability and effectiveness for the two use cases through two experiments. In the first experiment, we use SceneUI to create biased datasets where each bias is based on an object's colour or object type and the goal is to train a deep learning model that learns the biases. The results show that the model learns the biases well and thus, SceneUI can be used to control datasets which can be used to benchmark machine learning explainability methods. The second experiment generates a dataset of images using the OR query and training machine learning models on the dataset to ensure the suitability of the images for machine learning tasks. As the generated images will be used as ground truth given a query for a specialised machine learning model, the model is expected to identify the predicate objects in the image. The results show SceneUI can generate images based on objects in the query and can also accommodate the objects to have properties. The models identify important features in the SceneUI-generated images and are thus suitable to be used as ground truth. We also discuss the limitations and tradeoffs of SceneUI and present potential future directions for research to improve its scalability.

Commonsense knowledge (CK) in artificial intelligence (AI), is an expanding field of research. Because CK is intrinsically implicit, current datadriven machine learning models are still far from competent compared to humans in commonsense reasoning tasks. To minimize the gap between machine learning models with the goal of artificial general intelligence (AGI), researchers propose to collect such CK from human and text corpus for commonsense augmentation. Over the years there have been many ways that commonsense knowledge (in AI) has been implemented. However, there has not been a systematic review conducted on how CK can be organized, structured and categorized. The aim of this paper is to do a literature survey on how existing knowledge sources organize, structure and categorize within the general frameworks of CK. The organization can decide the design schema of a knowledge graph (KG), the structuring decides the format and the categorization decides what dimensions and criteria a KG employs.

Commonsense knowledge is the key of human intelligence in generalizing their knowledge to deal with complex tasks. Over the past years, a lot of research has been done in both natural language processing (NLP) and computer vision (CV) on leveraging commonsense knowledge to improve AI models. However, no systematic comparisons of existing work have been made between the two domains. Therefore this survey aims to provide an overview of how commonsense knowledge is used within NLP and CV and how research varies between these two domains and what future challenges it may hold. An observation made from this survey is that leveraging commonsense is more difficult in CV than NLP, as commonsense is mostly incorporated textually and datasets need to be filtered to make them more relevant for visual commonsense. We hope to promote further research and create a better understanding of commonsense knowledge and its applications with this survey.

Commonsense knowledge is information that all humans own and use to interpret common situations and react to them accordingly. This kind of information is necessary for the training of artificial intelligence models to reach a performance as close as possible to human performance. Researchers have developed methods that use crowdsourcing to collect this kind of knowledge from the general public. This research focuses on systematically surveying the existing literature about these methods. We created a taxonomy to describe and compare the existing work based on the following three measures that were the most common ones reported: efficiency, cost, and quality.

Common sense knowledge (CSK) comes naturally to humans, but is very hard for computers to comprehend. However it is critical for machines to behave intelligently, and as such collecting CSK has become a prevalent field of research. Whilst a lot of research has been done to develop CSK acquisition methods, not much work has been done to survey the literature that already exists. Furthermore the surveys that have been done are outdated, and as such there is a clear gap in the literature. This paper will survey the different approaches to CSK acquisition and evaluate their effectiveness, as a way of gauging their real life applicability. It will also compare the current \textit{state of the art} methods, to some previous work to illustrate the progress that has been made and project that into the future. Furthermore this paper will also create a taxonomy categorizing the surveyed literature in order give a better overview of existing methods. Finally, from the literature surveyed it is clear that these methods have made a lot of progress, but aren't quite yet at the same level as human performance. Nevertheless they have become robust enough to be deployed in real applications.

Spam calls are becoming an increasing problem, with people receiving multiple spam calls per month on average. Multiple Android applications exist that are able to detect spam calls and display a warning or block such calls. Little is known however on how these applications work and what numbers they block.
In this research, the following question is investigated: Can we do a brute force dynamic analysis on Android spam call blocking apps, to extract caller ID information from apps that cannot be or is not extracted through static analysis? A tool is created that is capable of doing such a dynamic analysis, by installing such an app on an emulator, sending emulated phone calls to the emulator, and using screenshot comparison techniques to determine whether the call is classified as allowed or blocked by the respective app. The tool developed in this research can, fully automated, test caller IDs on 8 different Android apps. Apart from a number of initial setup steps to install and configure the apps in the emulator, the tool takes about 1.5 seconds on average to analyze 1 caller IDs on one app.

The aim of this research is to provide a structured approach for dynamically analysing Android applications, focusing on applications that block or flag suspected spam caller IDs. This paper discusses ways to determine how an application stores the data regarding the phone numbers, and what APIs it utilizes. In order to develop a methodology of what to search for while performing dynamic analysis, the research in this paper focuses on analysing one such Android application, using it as a template. Additionally, it provides insights and limitations on this analysed application, and potential improvements both on the techniques and tools used.

Spam calls are a widespread problem as people receive about 14 spam calls per month on average. In response, tens of applications are available in Google’s Play Store that aim to block these calls. While these apps have hundreds of millions of installations, there’s a lack of research into these applications. Research has been done on the user experience of these apps, but the inner workings of the apps have not been explored. Analyzing the inner workings can provide useful insights to combat spam calls more effectively. In this paper, we analyze the elements of the AndroidManifest.xml file of a set of 10 spam call blocking applications to identify aspects that are related to the blocking of spam calls and to find commonalities among the apps. Occurrences of specific elements and their attributes are summarized in this paper. Permissions related to managing phone calls and accessing the internet were found to be common, and services were identified that screen and manage phone calls. 2 apps were found to allow other apps to use their phone number information database, but this analysis did not identify usages by other apps. A complete analysis of all available spam call blocking apps could show the general usage of these and other components.

Applications on Android phones which block spam calls could be argued a necessity for people unfortunate enough to have their number on spam lists. Third-party applications provide exten- sive, up-to-date blocklists to screen incoming calls. This paper analyses and describes how these types of applications store and access their data. First we analyse which applications keep their data offline on disk, then we take a look at applications which use the internet to query an in- coming phone call. We found four applications use a combination of both, two applications using exclusively online resources and four applications using offline resources only.

In order to combat increasing spam calls, many applications are developed and downloaded to block those calls. Some studies about performance of the applications were previously conducted, however, the actual processes the applications go through to intercept and block the spam calls are not well studied. This research presents a method to systematically explore Android APIs that are responsible for intercepting and blocking the spam calls.

Powerful predictive AI systems have demonstrated great potential in augmenting human decision-making. Recent empirical work has argued that the vision for optimal human-AI collaboration requires ‘appropriate reliance’ of humans on AI systems. However, accurately estimating the trustworthiness of AI advice at the instance level is quite challenging, especially in the absence of performance feedback pertaining to the AI system. In practice, the performance disparity of machine learning models on out-of-distribution data makes the dataset-specific performance feedback unreliable in human-AI collaboration. Inspired by existing literature on critical thinking and explanation-based human debugging, we propose the use of debugging an AI system as an intervention to foster appropriate reliance. In this paper, we explore whether a critical evaluation of AI performance within a debugging setting can better calibrate users’ assessment of an AI system and lead to more appropriate reliance. Through a quantitative empirical study (N = 234), we found that our proposed debugging intervention does not work as expected in facilitating appropriate reliance. Instead, we observe a decrease in reliance on the AI system after the intervention — potentially resulting from early exposure to the AI system’s weakness. We explored the dynamics of user confidence to help explain how inappropriate reliance patterns occur and found that human confidence is not independent of AI advice, which is potentially dangerous when trying to achieve appropriate reliance. Our findings have important implications for designing effective interventions to facilitate appropriate reliance and better human-AI collaboration

The state-of-the-art shows the potential of chatbots and other Machine Learning (ML) models to perform many tasks of high quality. Especially chatbots are already used by many companies to assist their customer service. However, chatbots will likely never be able to perform all tasks perfectly. Therefore, it is still the question whether such a chatbot is valuable for a business. Current research fails to describe how chatbots should be evaluated to compute the value of a chatbot for a business. In this research, we design an evaluation framework capturing the value of a chatbot in customer service. This framework consists of several key dimensions which should be computed in order to determine the value of the chatbot. To show that this evaluation framework captures the value of a chatbot, we perform a case study on water utility companies in The Netherlands. This case study showed the designed evaluation framework does capture the value of a chatbot in customer service.

The workflow of a data science practitioner includes gathering information from different sources and applying machine learning (ML) models. Such dispersed information can be combined through a process known as Data Integration (DI), which defines relations between entities and attributes. When all information is combined in one source suited for ML purposes, this source often contains duplicate data, resulting in longer operation times. Recent work has created algebraic rewrite rules for ML such that computations can be pushed down to individual sources. This method is referred to as factorized learning. In this work, we present an implementation of Amalur, a system for automated factorized learning that is novel because of its applicability in any DI scenario. Amalur shows significant speedups for some datasets and models, while for some other datasets and models, our implementation of factorized learning results in slowdowns. Whether factorized learning is efficient depends on the operations involved in the applied ML model and the data supplied to the model. The process of estimating the efficiency of factorization is known as cost estimation. Previous efforts on cost estimation do not generalize to the DI scenarios included in Amalur, are not tailored to different ML models, and are not properly evaluated. In this thesis, we create a cost estimation procedure for Amalur suitable for any DI scenario, which adapts to individual ML models. To evaluate this procedure we create a data generator capable of generating customizable DI scenarios. Our method outperforms its competitor on DI scenarios covered by the state of the art, and shows comparable performance on newly covered DI scenarios.

Current speed of data growth has exponentially increased over the past decade, highlighting the need of modern organizations for data discovery systems. Several (automated) schema matching approaches have been proposed to find related data, exploiting different parts of schema information (e.g. data type, data distribution, column name, etc.). However, research showed that single schema matching techniques fails to effectively match schemas, whilst combinatorial schema matching systems show more promise. With the introduction of combinatorial schema matching systems, new challenges arise regarding selection and combining strategies. This research attempts to explore different techniques for determining the importance of each matcher in a combinatorial schema matching system by determining the weights of each matcher and comparing them through a comprehensive evaluation.

Dataset discovery techniques originally required datasets to have the same domain which made them unsuitable to be used on a larger scale. To avoid this requirement, newer techniques use additional information, aside from the datasets being processed, to better understand the data. They might rely on a knowledge base that describes the meaning of data, or a lexical database that defines meaningful relations between the words contained within the data. The main problem with this approach is that these types of additional information have poor coverage of the data being analyzed.

I propose to use a type of information that I call dataset domain terms. These are terms, or data values, that represent the domain of a dataset. I provide a technique that can derive these dataset domain terms automatically from existing datasets which means they are easily available. The problem of poor coverage can also be mitigated by only discovering datasets for the domain represented by these dataset domain terms. I provide a dataset discovery technique that takes this approach with these dataset domain terms.

Through an evaluation, I show that these dataset domain terms are sufficiently representative of the domain to be used for dataset discovery. The accuracy of the dataset discovery technique is also shown to be comparable to state-of-the-art dataset discovery techniques, though its precision is lacking.

This makes it highly suitable to filter datasets before other dataset discovery techniques can be performed on them. The data from these filtered datasets should also have a limited range of domains. So subsequent dataset discovery techniques should be less affected by the poor coverage of the additional information they use to understand the data. This allows dataset discovery to be performed on a larger scale.

Recent works explain the DNN models that perform image classification tasks following the "attribution, human-in-the-loop, extraction" workflow. However, little work has looked into such an approach for explaining DNN models for language or multimodal tasks. To address this gap, we propose a framework that explains and assesses the model utilizing both the categorical/numerical features and the text while optimizing the "attribution, human-in-the-loop, extraction" workflow. In particular, our framework deals with limited human resources, especially when domain experts are required for human-in-the-loop tasks. It provides insight regarding which set of data should the human-in-the-loop tasks be brought in. We share the results of applying this framework to a multimodal transformer that performs text classification tasks for compliance detection in the financial context.