During the COVID-19 pandemic, the World Health Organization provided a checklist to help people distinguish between accurate and misinformation. In controlled experiments in the United States and Germany, we investigated the utility of this ordered checklist and designed an interactive version to lower the cost of acting on checklist items. Across interventions, we observe non-trivial differences in participants' performance in distinguishing accurate and misinformation between the two countries and discuss some possible reasons that may predict the future helpfulness of the checklist in different environments. The checklist item that provides source labels was most frequently followed and was considered most helpful. Based on our empirical findings, we recommend practitioners focus on providing source labels rather than interventions that support readers performing their own fact-checks, even though this recommendation may be influenced by the WHO's chosen order. We discuss the complexity of providing such source labels and provide design recommendations.
相關內容
Explainable AI (XAI) is a rapidly evolving field that aims to improve transparency and trustworthiness of AI systems to humans. One of the unsolved challenges in XAI is estimating the performance of these explanation methods for neural networks, which has resulted in numerous competing metrics with little to no indication of which one is to be preferred. In this paper, to identify the most reliable evaluation method in a given explainability context, we propose MetaQuantus -- a simple yet powerful framework that meta-evaluates two complementary performance characteristics of an evaluation method: its resilience to noise and reactivity to randomness. We demonstrate the effectiveness of our framework through a series of experiments, targeting various open questions in XAI, such as the selection of explanation methods and optimisation of hyperparameters of a given metric. We release our work under an open-source license to serve as a development tool for XAI researchers and Machine Learning (ML) practitioners to verify and benchmark newly constructed metrics (i.e., ``estimators'' of explanation quality). With this work, we provide clear and theoretically-grounded guidance for building reliable evaluation methods, thus facilitating standardisation and reproducibility in the field of XAI.
Large-scale generative models enabled the development of AI-powered code completion tools to assist programmers in writing code. However, much like other AI-powered tools, AI-powered code completions are not always accurate, potentially introducing bugs or even security vulnerabilities into code if not properly detected and corrected by a human programmer. One technique that has been proposed and implemented to help programmers identify potential errors is to highlight uncertain tokens. However, there have been no empirical studies exploring the effectiveness of this technique-- nor investigating the different and not-yet-agreed-upon notions of uncertainty in the context of generative models. We explore the question of whether conveying information about uncertainty enables programmers to more quickly and accurately produce code when collaborating with an AI-powered code completion tool, and if so, what measure of uncertainty best fits programmers' needs. Through a mixed-methods study with 30 programmers, we compare three conditions: providing the AI system's code completion alone, highlighting tokens with the lowest likelihood of being generated by the underlying generative model, and highlighting tokens with the highest predicted likelihood of being edited by a programmer. We find that highlighting tokens with the highest predicted likelihood of being edited leads to faster task completion and more targeted edits, and is subjectively preferred by study participants. In contrast, highlighting tokens according to their probability of being generated does not provide any benefit over the baseline with no highlighting. We further explore the design space of how to convey uncertainty in AI-powered code completion tools, and find that programmers prefer highlights that are granular, informative, interpretable, and not overwhelming.
Most works on the fairness of machine learning systems focus on the blind optimization of common fairness metrics, such as Demographic Parity and Equalized Odds. In this paper, we conduct a comparative study of several bias mitigation approaches to investigate their behaviors at a fine grain, the prediction level. Our objective is to characterize the differences between fair models obtained with different approaches. With comparable performances in fairness and accuracy, are the different bias mitigation approaches impacting a similar number of individuals? Do they mitigate bias in a similar way? Do they affect the same individuals when debiasing a model? Our findings show that bias mitigation approaches differ a lot in their strategies, both in the number of impacted individuals and the populations targeted. More surprisingly, we show these results even apply for several runs of the same mitigation approach. These findings raise questions about the limitations of the current group fairness metrics, as well as the arbitrariness, hence unfairness, of the whole debiasing process.
Mis- and disinformation are a substantial global threat to our security and safety. To cope with the scale of online misinformation, researchers have been working on automating fact-checking by retrieving and verifying against relevant evidence. However, despite many advances, a comprehensive evaluation of the possible attack vectors against such systems is still lacking. Particularly, the automated fact-verification process might be vulnerable to the exact disinformation campaigns it is trying to combat. In this work, we assume an adversary that automatically tampers with the online evidence in order to disrupt the fact-checking model via camouflaging the relevant evidence or planting a misleading one. We first propose an exploratory taxonomy that spans these two targets and the different threat model dimensions. Guided by this, we design and propose several potential attack methods. We show that it is possible to subtly modify claim-salient snippets in the evidence and generate diverse and claim-aligned evidence. Thus, we highly degrade the fact-checking performance under many different permutations of the taxonomy's dimensions. The attacks are also robust against post-hoc modifications of the claim. Our analysis further hints at potential limitations in models' inference when faced with contradicting evidence. We emphasize that these attacks can have harmful implications on the inspectable and human-in-the-loop usage scenarios of such models, and conclude by discussing challenges and directions for future defenses.
With an increased focus on incorporating fairness in machine learning models, it becomes imperative not only to assess and mitigate bias at each stage of the machine learning pipeline but also to understand the downstream impacts of bias across stages. Here we consider a general, but realistic, scenario in which a predictive model is learned from (potentially biased) training data, and model predictions are assessed post-hoc for fairness by some auditing method. We provide a theoretical analysis of how a specific form of data bias, differential sampling bias, propagates from the data stage to the prediction stage. Unlike prior work, we evaluate the downstream impacts of data biases quantitatively rather than qualitatively and prove theoretical guarantees for detection. Under reasonable assumptions, we quantify how the amount of bias in the model predictions varies as a function of the amount of differential sampling bias in the data, and at what point this bias becomes provably detectable by the auditor. Through experiments on two criminal justice datasets -- the well-known COMPAS dataset and historical data from NYPD's stop and frisk policy -- we demonstrate that the theoretical results hold in practice even when our assumptions are relaxed.
Finding a product online can be a challenging task for users. Faceted search interfaces, often in combination with recommenders, can support users in finding a product that fits their preferences. However, those preferences are not always equally weighted: some might be more important to a user than others (e.g. red is the favorite color, but blue is also fine) and sometimes preferences are even contradictory (e.g. the lowest price vs. the highest performance). Often, there is even no product that meets all preferences. In those cases, faceted search interfaces reach their limits. In our research, we investigate the potential of a search interface, which allows a preference-based ranking based on weighted search and facet terms. We performed a user study with 24 participants and measured user satisfaction and system performance. The results show that with the preference-based search interface the users were given more alternatives that best meet their preferences and that they are more satisfied with the selected product than with a search interface using standard facets. Furthermore, in this work we study the relationship between user satisfaction and search precision within the whole search session and found first indications that there might be a relation between them.
For domains in which a recommender provides repeated content suggestions, agent preferences may evolve over time as a function of prior recommendations, and algorithms must take this into account for long-run optimization. Recently, Agarwal and Brown (2022) introduced a model for studying recommendations when agents' preferences are adaptive, and gave a series of results for the case when agent preferences depend {\it uniformly} on their history of past selections. Here, the recommender shows a $k$-item menu (out of $n$) to the agent at each round, who selects one of the $k$ items via their history-dependent {\it preference model}, yielding a per-item adversarial reward for the recommender. We expand this setting to {\it non-uniform} preferences, and give a series of results for {\it $\gamma$-discounted} histories. For this problem, the feasible regret benchmarks can depend drastically on varying conditions. In the ``large $\gamma$'' regime, we show that the previously considered benchmark, the ``EIRD set'', is attainable for any {\it smooth} model, relaxing the ``local learnability'' requirement from the uniform memory case. We introduce ``pseudo-increasing'' preference models, for which we give an algorithm which can compete against any item distribution with small uniform noise (the ``smoothed simplex''). We show NP-hardness results for larger regret benchmarks in each case. We give another algorithm for pseudo-increasing models (under a restriction on the adversarial nature of the reward functions), which works for any $\gamma$ and is faster when $\gamma$ is sufficiently small, and we show a super-polynomial regret lower bound with respect to EIRD for general models in the ``small $\gamma$'' regime. We conclude with a pair of algorithms for the memoryless case.
Background. The mass transition to remote work amid the COVID-19 pandemic profoundly affected software professionals, who abruptly shifted into ostensibly temporary home offices. The effects of this transition on these professionals are complex, depending on the particularities of the context and individuals. Recent studies advocate for remote structures to create opportunities for many equity-deserving groups; however, remote work can also be challenging for some individuals, such as women and individuals with disabilities. Objective. This study aims to investigate the effects of remote work on LGBTQIA+ software professionals. Method. Grounded theory methodology was applied based on information collected from two main sources: a survey questionnaire with a sample of 57 LGBTQIA+ software professionals and nine follow-up interviews with individuals from this sample. This sample included professionals of different genders, ethnicities, sexual orientations, and levels of experience. Findings. Our findings demonstrate that (1) remote work benefits LGBTQIA+ people by increasing security and visibility; (2) remote work harms LGBTQIA+ software professionals through isolation and invisibility; (3) the benefits outweigh the drawbacks; (4) the drawbacks can be mitigated by supportive measures developed by software companies. Conclusion. This paper investigated how remote work can affect LGBTQIA+ software professionals and presented a set of recommendations on how software companies can address the benefits and limitations associated with this work model. In summary, we concluded that remote work is crucial in increasing diversity and inclusion in the software industry.
This paper focuses on the expected difference in borrower's repayment when there is a change in the lender's credit decisions. Classical estimators overlook the confounding effects and hence the estimation error can be magnificent. As such, we propose another approach to construct the estimators such that the error can be greatly reduced. The proposed estimators are shown to be unbiased, consistent, and robust through a combination of theoretical analysis and numerical testing. Moreover, we compare the power of estimating the causal quantities between the classical estimators and the proposed estimators. The comparison is tested across a wide range of models, including linear regression models, tree-based models, and neural network-based models, under different simulated datasets that exhibit different levels of causality, different degrees of nonlinearity, and different distributional properties. Most importantly, we apply our approaches to a large observational dataset provided by a global technology firm that operates in both the e-commerce and the lending business. We find that the relative reduction of estimation error is strikingly substantial if the causal effects are accounted for correctly.
Although measuring held-out accuracy has been the primary approach to evaluate generalization, it often overestimates the performance of NLP models, while alternative approaches for evaluating models either focus on individual tasks or on specific behaviors. Inspired by principles of behavioral testing in software engineering, we introduce CheckList, a task-agnostic methodology for testing NLP models. CheckList includes a matrix of general linguistic capabilities and test types that facilitate comprehensive test ideation, as well as a software tool to generate a large and diverse number of test cases quickly. We illustrate the utility of CheckList with tests for three tasks, identifying critical failures in both commercial and state-of-art models. In a user study, a team responsible for a commercial sentiment analysis model found new and actionable bugs in an extensively tested model. In another user study, NLP practitioners with CheckList created twice as many tests, and found almost three times as many bugs as users without it.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191