In this paper, we argue that the prevailing approach to training and evaluating machine learning models often fails to consider their real-world application within organizational or societal contexts, where they are intended to create beneficial value for people. We propose a shift in perspective, redefining model assessment and selection to emphasize integration into workflows that combine machine predictions with human expertise, particularly in scenarios requiring human intervention for low-confidence predictions. Traditional metrics like accuracy and f-score fail to capture the beneficial value of models in such hybrid settings. To address this, we introduce a simple yet theoretically sound “value” metric that incorporates task-specific costs for correct predictions, errors, and rejections, offering a practical framework for real-world evaluation. Through extensive experiments, we show that existing metrics fail to capture real-world needs, often leading to suboptimal choices in terms of value when used to rank classifiers. Furthermore, we emphasize the critical role of calibration in determining model value, showing that simple, well-calibrated models can often outperform more complex models that are challenging to calibrate.

Despite the impressive performance of Artificial Intelligence (AI) systems, their robustness remains elusive and constitutes a key issue that impedes large-scale adoption. Besides, robustness is interpreted differently across domains and contexts of AI. In this work, we systematically survey recent progress to provide a reconciled terminology of concepts around AI robustness. We introduce three taxonomies to organize and describe the literature both from a fundamental and applied point of view: (1) methods and approaches that address robustness in different phases of the machine learning pipeline; (2) methods improving robustness in specific model architectures, tasks, and systems; and in addition, (3) methodologies and insights around evaluating the robustness of AI systems, particularly the tradeoffs with other trustworthiness properties. Finally, we identify and discuss research gaps and opportunities and give an outlook on the field. We highlight the central role of humans in evaluating and enhancing AI robustness, considering the necessary knowledge they can provide, and discuss the need for better understanding practices and developing supportive tools in the future.

Due to the significant time and effort required for handcrafting translations, most manga never leave the domestic Japanese market. Automatic manga translation is a promising potential solution. However, it is a budding and underdeveloped field and presents complexities even greater than those found in standard translation due to the need to effectively incorporate visual elements into the translation process to resolve ambiguities. In this work, we investigate to what extent multimodal large language models (LLMs) can provide effective manga translation, thereby assisting manga authors and publishers in reaching wider audiences. Specifically, we propose a methodology that leverages the vision component of multimodal LLMs to improve translation quality and evaluate the impact of translation unit size, context length, and propose a token efficient approach for manga translation. Moreover, we introduce a new evaluation dataset – the first parallel Japanese-Polish manga translation dataset – as part of a benchmark to be used in future research. Finally, we contribute an open-source software suite, enabling others to benchmark LLMs for manga translation. Our findings demonstrate that our proposed methods achieve state-of-the-art results for Japanese-English translation and set a new standard for Japanese-Polish.

Most existing bundle generation approaches fall short in generating fixed-size bundles. Furthermore, they often neglect the underlying user intents reflected by the bundles in the generation process, resulting in less intelligible bundles. This paper addresses these limitations through the exploration of two interrelated tasks, i.e., personalized bundle generation and the underlying intent inference, based on different user sessions. Inspired by the reasoning capabilities of large language models (LLMs), we propose an adaptive in-context learning paradigm, which allows LLMs to draw tailored lessons from related sessions as demonstrations, enhancing the performance on target sessions. Specifically, we first employ retrieval augmented generation to identify nearest neighbor sessions, and then carefully design prompts to guide LLMs in executing both tasks on these neighbor sessions. To tackle reliability and hallucination challenges, we further introduce (1) a self-correction strategy promoting mutual improvements of the two tasks without supervision signals and (2) an auto-feedback mechanism for adaptive supervision based on the distinct mistakes made by LLMs on different neighbor sessions. Thereby, the target session can gain customized lessons for improved performance by observing the demonstrations of its neighbor sessions. Experiments on three real-world datasets demonstrate the effectiveness of our proposed method.

Relation extraction methods are currently dominated by deep neural models, which capture complex statistical patterns while being brittle and vulnerable to perturbations in data and distribution. Explainability techniques offer a means for understanding such vulnerabilities, and thus represent an opportunity to mitigate future errors; yet, existing methods are limited to describing what the model 'knows', while totally failing at explaining what the model does not know. This paper presents a new method for diagnosing model predictions and detecting potential inaccuracies. Our approach involves breaking down the problem into two components: (i) determining the necessary knowledge the model should possess for accurate prediction, through human annotations, and (ii) assessing the actual knowledge possessed by the model, using explainable AI methods (XAI). We apply our method to several relation extraction tasks and conduct an empirical study leveraging human specifications of what a model should know and does not know. Results show that human workers are capable of accurately specifying the model should-knows, despite variations in the specification, that the alignment between what a model really knows and what it should know is indeed indicative of model accuracy, and that the unknowns identified through our methods allow to foresee future errors that may or may not have been observed otherwise.

Large Language Models (LLMs) are expected to significantly impact various socio-technical systems, offering transformative possibilities for improved interaction between humans and technology. However, their integration poses complex challenges due to the intricate interplay between societal structures, human behaviour, and technological innovation. This research explores these multifaceted challenges, emphasising the need for a human-centered approach in integrating LLMs to ensure that technological advancements are aligned with ethical standards and societal needs. Utilizing a structured methodology comprising a workshop, literature analysis, and expert collaborations, the study uses a multi-dimensional human-centered AI framework to guide the responsible integration of LLMs. Key insights include the importance of inclusive data, considering unintended consequences, maintaining privacy, and respecting intellectual property rights. The paper identifies and advocates for principles like human-in-the-loop, continuous longitudinal studies, proactive awareness campaigns, and regular audits to develop LLMs that are ethically sound, adaptable, and effectively integrated into various socio-technical systems, thus addressing user needs and broader societal impacts. The paper also underlines the importance of collaboration among academia, industry, and policymakers to develop LLMs that are ethically aligned, socially beneficial, and adaptable to future societal needs. The findings offer valuable insights into the strategic integration of LLMs, advocating for a broader research perspective beyond industrial motivations to fully understand and leverage LLMs in socio-technical landscapes.

Clinicians increasingly pay attention to Artificial Intelligence (AI) to improve the quality and timeliness of their services. There are converging opinions on the need for Explainable AI (XAI) in healthcare. However, prior work considers explanations as stationary entities with no account for the temporal dynamics of patient care. In this work, we involve 16 Idiopathic Pulmonary Fibrosis (IPF) clinicians from a European university medical centre and investigate their evolving uses and purposes for explainability throughout patient care. By applying a patient journey map for IPF, we elucidate clinicians' informational needs, how human agency and patient-specific conditions can influence the interaction with XAI systems, and the content, delivery, and relevance of explanations over time. We discuss implications for integrating XAI in clinical contexts and more broadly how explainability is defined and evaluated. Furthermore, we reflect on the role of medical education in addressing epistemic challenges related to AI literacy.

Federated Learning (FL) is an important privacy-preserving learning paradigm that plays an important role in the Intelligent Internet of Things. Training a global model in FL, however, is vulnerable to the data noise across the clients. In this paper, we introduce FedTrans, a novel client-transparent client utility estimation method designed to guide client selection for noisy scenarios, mitigating performance degradation problems. To estimate the client utility, we propose a Bayesian framework that models client utility and its relationships with the weight parameters and the performance of local models. We then introduce a variational inference algorithm to effectively infer client utility at the FL server, given only a small amount of auxiliary data. Our evaluation results demonstrate that leveraging FedTrans to select the clients can improve the accuracy performance (up to 7.8%), ensuring the robustness of FL in noisy scenarios.

Concepts are an important construct in semantics, based on which humans understand the world with various levels of abstraction. With the recent advances in explainable artificial intelligence (XAI), concept-level explanations are receiving an increasing amount of attention from the broad research community. However, laypeople may find such explanations difficult to digest due to the potential knowledge gap and the concomitant cognitive load. Inspired by prior work that has explored analogies and sensemaking, we argue that augmenting concept-level explanations with analogical inference information from commonsense knowledge can be a potential solution to tackle this issue. To investigate the validity of our proposition, we first designed an effective analogy-based explanation generation method and collected 600 analogy-based explanations from 100 crowd workers. Next, we proposed a set of structured dimensions for the qualitative assessment of such explanations, and conducted an empirical evaluation of the generated analogies with experts. Our findings revealed significant positive correlations between the qualitative dimensions of analogies and the perceived helpfulness of analogy-based explanations, suggesting the effectiveness of the dimensions. To understand the practical utility and the effectiveness of analogybased explanations in assisting human decision-making, we conducted a follow-up empirical study (N = 280) on a skin cancer detection task with non-expert humans and an imperfect AI system. Thus, we designed a between-subjects study spanning five different experimental conditions with varying types of explanations. The results of our study confirmed that a knowledge gap can prevent participants from understanding concept-level explanations. Consequently, when only the target domain of our designed analogy-based explanation was provided (in a specific experimental condition), participants demonstrated relatively more appropriate reliance on the AI system. In contrast to our expectations, we found that analogies were not effective in fostering appropriate reliance. We carried out a qualitative analysis of the open-ended responses from participants in the study regarding their perceived usefulness of explanations and analogies. Our findings suggest that human intuition and the perceived plausibility of analogies may have played a role in affecting user reliance on the AI system. We also found that the understanding of commonsense explanations varied with the varying experience of the recipient user, which points out the need for further work on personalization when leveraging commonsense explanations. In summary, although we did not find quantitative support for our hypotheses around the benefits of using analogies, we found considerable qualitative evidence suggesting the potential of high-quality analogies in aiding non-expert users in their decision making with AI-assistance. These insights can inform the design of future methods for the generation and use of effective analogy-based explanations.

Modern Knowledge Graphs (KGs) are inevitably noisy due to the nature of their construction process. Existing robust learning techniques for noisy KGs mostly focus on triple facts, where the factwise confidence is straightforward to evaluate. However, hyperrelational facts, where an arbitrary number of key-value pairs are associated with a base triplet, have become increasingly popular in modern KGs, but significantly complicate the confidence assessment of the fact. Against this background, we study the problem of robust link prediction over noisy hyper-relational KGs, and propose NYLON, a Noise-resistant hYper-reLatiONal link prediction technique via active crowd learning. Specifically, beyond the traditional fact-wise confidence, we first introduce element-wise confidence measuring the fine-grained confidence of each entity or relation of a hyper-relational fact. We connect the element- and fact-wise confidences via a “least confidence” principle to allow efficient crowd labeling. NYLON is then designed to systematically integrate three key components, where a hyper-relational link predictor uses the fact-wise confidence for robust prediction, a cross-grained confidence evaluator predicts both element- and fact-wise confidences, and an effort-efficient active labeler selects informative facts for crowd annotators to label using an efficient labeling mechanism guided by the element-wise confidence under the “least confidence” principle and further followed by data augmentation. We evaluate NYLON on three real-world KG datasets against a sizeable collection of baselines. Results show that NYLON achieves superior and robust performance in both link prediction and error detection tasks on noisy KGs, and outperforms best baselines by 2.42-10.93% and 3.46-10.65% in the two tasks, respectively.

Textbook question answering is challenging as it aims to automatically answer various questions on textbook lessons with long text and complex diagrams, requiring reasoning across modalities. In this work, we propose MRHF, a novel framework that incorporates dense passage re-ranking and the mixture-of-experts architecture for TQA. MRHF proposes a novel query augmentation method for diagram questions and then adopts multi-stage dense passage re-ranking with large pretrained retrievers for retrieving paragraph-level contexts. Then it employs a unified question solver to process different types of text questions. Considering the rich blobs and relation knowledge contained in diagrams, we propose to perform multimodal feature fusion over the retrieved context and the heterogeneous diagram features. Furthermore, we introduce the mixture-of-experts architecture to solve the diagram questions to learn from both the rich text context and the complex diagrams and mitigate the possible negative effects between features of the two modalities. We test the framework on the CK12-TQA benchmark dataset, and the results show that MRHF outperforms the state-of-the-art results in all types of questions. The ablation and case study also demonstrates the effectiveness of each component of the framework.

Visible light positioning (VLP) based on the received signal strength (RSS) can leverage a dense deployment of LEDs in future lighting infrastructure to provide accurate and energy-efficient indoor positioning. However, its positioning accuracy heavily depends on the density of collected fingerprints, which is labor-intensive. In this work, we propose a data pre-processing method, including data cleaning and data augmentation, to construct reliable and dense fingerprint samples, thereby alleviating the impact of noisy samples as well as reducing labor intensity. Extensive experiments demonstrate that our proposed method achieves an average positioning error of 1.7 cm, utilizing a sparse dataset that reduces the fingerprint collection effort by 98 percent. Running a tinyML-based model for VLP on the Arduino Nano microcontroller, we also show the possibilities for deploying RSS fingerprint-based VLP systems on resource-constrained embedded devices for real-world applications.

This Special Issue of the Journal of Data and Information Quality (JDIQ) contains novel theoretical and methodological contributions on data curation involving humans in the loop. In this editorial, we summarize the scope of the issue and briefly describe its content.

Handling failures in computer vision systems that rely on deep learning models remains a challenge. While an increasing number of methods for bug identification and correction are proposed, little is known about how practitioners actually search for failures in these models. We perform an empirical study to understand the goals and needs of practitioners, the workflows and artifacts they use, and the challenges and limitations in their process. We interview 18 practitioners by probing them with a carefully crafted failure handling scenario. We observe that there is a great diversity of failure handling workflows in which cooperations are often necessary, that practitioners overlook certain types of failures and bugs, and that they generally do not rely on potentially relevant approaches and tools originally stemming from research. These insights allow to draw a list of research opportunities, such as creating a library of best practices and more representative formalisations of practitioners' goals, developing interfaces to exploit failure handling artifacts, as well as providing specialized training.

In many practical applications, machine learning models are embedded into a pipeline involving a human actor that decides whether to trust the machine prediction or take a default route (e.g., classify the example herself). Selective classifiers have the option to abstain from making a prediction on an example they do not feel confident about. Recently, the notion of the value of a machine learning model has been introduced as a way to jointly consider the benefit of a correct prediction, the cost of an error, and that of abstaining. In this paper, we study how active learning of selective classifiers is affected by the focus on value. We show that the performance of the state-of-the-art active learning strategies drops significantly when we evaluate them based on value rather than accuracy. Finally, we propose a novel value-aware active learning strategy that outperforms the state-of-the-art ones when the cost of incorrect predictions substantially outweighs that of abstaining.

Federated Learning (FL) is an important privacy-preserving learning paradigm that is expected to play an essential role in the future Intelligent Internet of Things (IoT). However, model training in FL is vulnerable to noise and the statistical heterogeneity of local data across IoT clients. In this paper, we propose FedNaWi, a “Go Narrow, Then Wide” client selection method that speeds up the FL training, achieves higher model performance, while requiring no additional data or sensitive information transfer from clients. Our method first selects reliable clients (i.e., going narrow) which allows the global model to quickly improve its performance and then includes less reliable clients (i.e., going wide) to exploit more IoT data of clients to further improve the global model. To profile client utility, we introduce a unified Bayesian framework to model the client utility at the FL server, assisted by a small amount of auxiliary data. We conduct extensive evaluations with 5 state-of-the-art FL methods, on 3 IoT tasks and under 7 different types of label and feature noise. We build an FL testbed with 38 IoT nodes (20 nodes run on Raspberry Pi 4B and 18 nodes run on Jetson Nano) for the evaluation. Our results show that FedNaWi improves the FL accuracy substantially and significantly reduces energy consumption. In particular, FedNaWi improves the accuracy from 35% to 75% in the non-IID Dirichlet setting, and reduces the average energy consumption by 55%.

Hate speech moderation remains a challenging task for social media platforms. Human-AI collaborative systems offer the potential to combine the strengths of humans' reliability and the scalability of machine learning to tackle this issue effectively. While methods for task handover in human-AI collaboration exist that consider the costs of incorrect predictions, insufficient attention has been paid to accurately estimating these costs. In this work, we propose a value-sensitive rejection mechanism that automatically rejects machine decisions for human moderation based on users' value perceptions regarding machine decisions. We conduct a crowdsourced survey study with 160 participants to evaluate their perception of correct and incorrect machine decisions in the domain of hate speech detection, as well as occurrences where the system rejects making a prediction. Here, we introduce Magnitude Estimation, an unbounded scale, as the preferred method for measuring user (dis)agreement with machine decisions. Our results show that Magnitude Estimation can provide a reliable measurement of participants' perception of machine decisions. By integrating user-perceived value into human-AI collaboration, we further show that it can guide us in 1) determining when to accept or reject machine decisions to obtain the optimal total value a model can deliver and 2) selecting better classification models as compared to the more widely used target of model accuracy.

Recently, one critical issue looms large in the field of recommender systems - there are no effective benchmarks for rigorous evaluation - which consequently leads to unreproducible evaluation and unfair comparison. We, therefore, conduct studies from the perspectives of practical theory and experiments, aiming at benchmarking recommendation for rigorous evaluation. Regarding the theoretical study, a series of hyper-factors affecting recommendation performance throughout the whole evaluation chain are systematically summarized and analyzed via an exhaustive review on 141 papers published at eight top-tier conferences within 2017-2020. We then classify them into model-independent and model-dependent hyper-factors, and different modes of rigorous evaluation are defined and discussed in-depth accordingly. For the experimental study, we release DaisyRec 2.0 library by integrating these hyper-factors to perform rigorous evaluation, whereby a holistic empirical study is conducted to unveil the impacts of different hyper-factors on recommendation performance. Supported by the theoretical and experimental studies, we finally create benchmarks for rigorous evaluation by proposing standardized procedures and providing performance of ten state-of-the-arts across six evaluation metrics on six datasets as a reference for later study. Overall, our work sheds light on the issues in recommendation evaluation, provides potential solutions for rigorous evaluation, and lays foundation for further investigation.

Fairness toolkits are developed to support machine learning (ML) practitioners in using algorithmic fairness metrics and mitigation methods. Past studies have investigated practical challenges for toolkit usage, which are crucial to understanding how to support practitioners. However, the extent to which fairness toolkits impact practitioners’ practices and enable reflexivity around algorithmic harms remains unclear (i.e., distributive unfairness beyond algorithmic fairness, and harms that are not related to the outputs of ML systems). Little is currently understood about the root factors that fragment practices when using fairness toolkits and how practitioners reflect on algorithmic harms. Yet, a deeper understanding of these facets is essential to enable the design of support tools for practitioners. To investigate the impact of toolkits on practices and identify factors that shape these practices, we carried out a qualitative study with 30 ML practitioners with varying backgrounds. Through a mixed within and between-subjects design, we tasked the practitioners with developing an ML model, and analyzed their reported practices to surface potential factors that lead to differences in practices. Interestingly, we found that fairness toolkits act as double-edge swords — with potentially positive and negative impacts on practices. Our findings showcase a plethora of human and organizational factors that play a key role in the way toolkits are envisioned and employed. These results bear implications for the design of future toolkits and educational training for practitioners and call for the creation of new policies to handle the organizational constraints faced by practitioners.

Evaluating design ideas is necessary to predict their success and assess their impact early on in the process. Existing methods rely either on metrics computed by systems that are effective but subject to errors and bias, or experts' ratings, which are accurate but expensive and long to collect. Crowdsourcing offers a compelling way to evaluate a large number of design ideas in a short amount of time while being cost-effective. Workers' evaluation is, however, less reliable and might substantially differ from experts' evaluation. In this work, we investigate workers' rating behavior and compare it with experts. First, we instrument a crowdsourcing study where we asked workers to evaluate design ideas from three innovation challenges. We show that workers share similar insights with experts but tend to rate more generously and weigh certain criteria more importantly. Next, we develop a hybrid human-AI approach that combines a machine learning model with crowdsourcing to evaluate ideas. Our approach models workers' reliability and bias while leveraging ideas' textual content to train a machine learning model. It is able to incorporate experts' ratings whenever available, to supervise the model training and infer worker performance. Results show that our framework outperforms baseline methods and requires significantly less training data from experts, thus providing a viable solution for evaluating ideas at scale.