Academic literature on the topic 'Bias mitigation techniques'

Abstract Since the 1960s and Wason’s famous experiments, the research into confirmation bias has not only become more field specific and realistic but it has also gradually shifted towards bias mitigation rather than studying the phenomenon of confirmation bias per se. This has become clear not least when it comes to confirmation bias in criminal investigations and proceedings. The chapter summarizes manifestations as well as possible debiasing techniques for specific settings during investigative and litigation phases. This includes interviews, suspect-line ups, crime scene investigations, forensic analysis, and so on (investigative phase) as well as prosecutors’ and judges’ decision-making (litigation phase). It is also clear that mitigating confirmation bias in criminal cases in a wider sense will require joint efforts of all actors included, and is to a certain extent a question of legal policy. It is essential that adjustments are made with a firm basis in research. In this chapter, important suggestions for future research, primarily focusing on debiasing techniques, are outlined.

Machine learning models have been instrumental in making decisions across domains, like mortgage lending and risk assessment in finance. However, these models have been found susceptible to biases, causing unfair decisions for a specific group of individuals. Such bias is generally based on some protected (or sensitive) attributes, such as age, sex, or race, and is still prevalent due to historical context or algorithmic bias. There have been several efforts to ensure equal opportunities for each individual/group, based on creditworthiness, rather than any social bias. Several pre-, in- and post-processing bias mitigation techniques have been proposed. However, these techniques perform data transformation or design new constraint/cost functions, which are task-specific, to achieve a fair prediction. Such techniques even require further access to the complete training/testing data. This paper proposes a novel post-processing bias mitigation technique that employs a model interpretation strategy to find the responsible model weights causing the bias. Pruning only a few model weights exhibits group fairness in model predictions while maintaining competitive accuracy levels, thus aligning with the goals of fairness and efficiency in decision-making. The proposed scheme requires access to only a few data samples representing the protected attributes, without exposing the complete training data. Through extensive experiments with multiple census datasets/methods, we demonstrate the efficacy of our approach, achieving up to a significant 50% reduction in bias while preserving the overall accuracy.

Artificial Intelligence (AI) offers great promise for healthcare, but integrating it comes with challenges. Over-reliance on AI systems can lead to automation bias, necessitating human oversight. Ethical considerations, transparency, and collaboration between healthcare providers and AI developers are crucial. Pursuing ethical frameworks, bias mitigation techniques, and transparency measures is key to advancing AI’s role in healthcare while upholding patient safety and quality care.

Abstract The importance of the so-called human factor for the outcome and accuracy of criminal law and procedure has been acknowledged for a long time but only recently has it been specified and more directly addressed in research. Confirmation bias specifically has been identified as a particularly problematic aspect of human reasoning that impacts on legal actors and criminal practitioners subconsciously. This means that even judges or prosecutors who are expected to be objective, and who may perceive of themselves as objective, may display this form of bias. Following the subconscious nature of confirmation bias, legal actors and criminal practitioners are unlikely to detect and mitigate the bias in themselves. Instead, effective bias mitigation requires scientific evaluations by researchers using established methodologies to evaluate different debiasing techniques.

Developing and delivering a project to an agreed schedule is fundamentally what project managers do. There is still an ongoing debate about schedule delays. This research investigates the development of schedules through semi-structured in-depth interviews. The findings reveal that half of the respondents believe that delays reported in the media are not real and should be attributed to scope changes. IT project managers estimating techniques include bottom-up estimates, analogy, and expert judgement. Impeding factors reported for the development of realistic schedules were technical (e.g. honest mistakes) and political (e.g. completion dates imposed by the sponsor). Respondents did not mention any psychological factors, although most were aware of optimism bias. However, they were not familiar with approaches to mitigate its impacts. Yet, when these techniques were mentioned, the overwhelming majority agreed that these mitigation approaches would change their schedule estimate.

This chapter explores strategies for enhancing the performance of AI-powered humanoid robots through innovation. It begins with an overview of the current landscape of humanoid robots, highlighting their diverse applications across industries. The review examines existing performance metrics and identifies areas for improvement. Subsequent sections delve into specific avenues for innovation, including advancements in cognitive capabilities, motor skills, emotional intelligence, and human-robot interaction. Leveraging machine learning techniques for continuous improvement is also explored. Ethical considerations, such as privacy concerns and bias mitigation, are addressed, along with challenges associated with societal impact. The chapter concludes with case studies showcasing successful implementations of performance-enhancing strategies and outlines potential future directions for research and development in the field.

The integration of artificial intelligence (AI) in prosthetic technology has led to significant advancements in adaptive control, improving functionality and user experience. Among AI-driven approaches, the combination of unsupervised learning and reinforcement learning has emerged as a transformative solution for real-time prosthetic adaptation. Unsupervised learning enables prosthetic devices to extract latent patterns from high-dimensional sensory data, allowing for autonomous identification of movement dynamics. Reinforcement learning, on the other hand, refines prosthetic control strategies through continuous reward-driven optimization, ensuring personalized adaptability without extensive manual calibration. This chapter explores the synergy between these learning paradigms, addressing key challenges such as sample efficiency, real-time implementation, and computational constraints. The impact of clustering techniques, dimensionality reduction, and generative models in enhancing prosthetic adaptability is analyzed. Ethical concerns, including data privacy, bias mitigation, and transparency, are also examined to ensure responsible AI deployment in prosthetic applications. The discussion highlights future research directions, emphasizing the need for efficient model generalization, lightweight AI architectures, and cloud-integrated learning frameworks to advance the next generation of intelligent prosthetic systems.

Artificial Intelligence (AI) holds great promise for healthcare, promising improved patient outcomes and streamlining processes. Nevertheless, this transformational journey comes with numerous potential pitfalls that warrant attention. This comprehensive review explores some key challenges involved with integrating AI into medicine. First and foremost is the risk of over-reliance on AI systems. Users often rely on recommendations provided by AI to follow without question, potentially causing automation bias. Human oversight is essential to avoid mistakes and patient harm; failure to provide such oversight could have serious repercussions that necessitate having someone in control at all times - emphasizing the necessity for having a human-in-the-loop approach. Ethical considerations must always come first when developing AI systems, with privacy, informed consent, and data protection as non-negotiable obligations for patients and organizations. Transparency and accountability within AI systems are necessary to quickly identify biases or errors to enable AI development with integrity that mitigates bias, ensures fairness, and maintains transparency. Ethical AI development involves ongoing efforts made with great diligence by developers to mitigate any bias, ensure fairness, and maintain transparency. These principles form the bedrock upon which ethical development depends. Collaboration between healthcare providers and AI developers is of utmost importance for patient safety and well-being; healthcare providers must protect patient data while developers must ensure AI systems adhere to legal and ethical requirements. AI and healthcare present significant challenges. Ethical frameworks, bias mitigation techniques, and transparency measures must all be pursued to advance AI’s role within healthcare delivery systems. We can unleash AI’s full potential by overcoming such hurdles while upholding patient safety, ethics, and quality care as the cornerstones of healthcare innovation.

A significant challenge in detecting and mitigating bias is creating a mindset amongst AI developers to address unfairness. The current literature on fairness is broad, and the learning curve to distinguish where to use existing metrics and techniques for bias detection or mitigation is difficult. This survey systematises the state-of-the-art about distinct notions of fairness and relative techniques for bias mitigation according to the AI lifecycle. Gaps and challenges identified during the development of this work are also discussed.

Objectives. To examine the evidence on whether and how healthcare algorithms (including algorithm-informed decision tools) exacerbate, perpetuate, or reduce racial and ethnic disparities in access to healthcare, quality of care, and health outcomes, and examine strategies that mitigate racial and ethnic bias in the development and use of algorithms. Data sources. We searched published and grey literature for relevant studies published between January 2011 and February 2023. Based on expert guidance, we determined that earlier articles are unlikely to reflect current algorithms. We also hand-searched reference lists of relevant studies and reviewed suggestions from experts and stakeholders. Review methods. Searches identified 11,500 unique records. Using predefined criteria and dual review, we screened and selected studies to assess one or both Key Questions (KQs): (1) the effect of algorithms on racial and ethnic disparities in health and healthcare outcomes and (2) the effect of strategies or approaches to mitigate racial and ethnic bias in the development, validation, dissemination, and implementation of algorithms. Outcomes of interest included access to healthcare, quality of care, and health outcomes. We assessed studies’ methodologic risk of bias (ROB) using the ROBINS-I tool and piloted an appraisal supplement to assess racial and ethnic equity-related ROB. We completed a narrative synthesis and cataloged study characteristics and outcome data. We also examined four Contextual Questions (CQs) designed to explore the context and capture insights on practical aspects of potential algorithmic bias. CQ 1 examines the problem’s scope within healthcare. CQ 2 describes recently emerging standards and guidance on how racial and ethnic bias can be prevented or mitigated during algorithm development and deployment. CQ 3 explores stakeholder awareness and perspectives about the interaction of algorithms and racial and ethnic disparities in health and healthcare. We addressed these CQs through supplemental literature reviews and conversations with experts and key stakeholders. For CQ 4, we conducted an in-depth analysis of a sample of six algorithms that have not been widely evaluated before in the published literature to better understand how their design and implementation might contribute to disparities. Results. Fifty-eight studies met inclusion criteria, of which three were included for both KQs. One study was a randomized controlled trial, and all others used cohort, pre-post, or modeling approaches. The studies included numerous types of clinical assessments: need for intensive care or high-risk care management; measurement of kidney or lung function; suitability for kidney or lung transplant; risk of cardiovascular disease, stroke, lung cancer, prostate cancer, postpartum depression, or opioid misuse; and warfarin dosing. We found evidence suggesting that algorithms may: (a) reduce disparities (i.e., revised Kidney Allocation System, prostate cancer screening tools); (b) perpetuate or exacerbate disparities (e.g., estimated glomerular filtration rate [eGFR] for kidney function measurement, cardiovascular disease risk assessments); and/or (c) have no effect on racial or ethnic disparities. Algorithms for which mitigation strategies were identified are included in KQ 2. We identified six types of strategies often used to mitigate the potential of algorithms to contribute to disparities: removing an input variable; replacing a variable; adding one or more variables; changing or diversifying the racial and ethnic composition of the patient population used to train or validate a model; creating separate algorithms or thresholds for different populations; and modifying the statistical or analytic techniques used by an algorithm. Most mitigation efforts improved proximal outcomes (e.g., algorithmic calibration) for targeted populations, but it is more challenging to infer or extrapolate effects on longer term outcomes, such as racial and ethnic disparities. The scope of racial and ethnic bias related to algorithms and their application is difficult to quantify, but it clearly extends across the spectrum of medicine. Regulatory, professional, and corporate stakeholders are undertaking numerous efforts to develop standards for algorithms, often emphasizing the need for transparency, accountability, and representativeness. Conclusions. Algorithms have been shown to potentially perpetuate, exacerbate, and sometimes reduce racial and ethnic disparities. Disparities were reduced when race and ethnicity were incorporated into an algorithm to intentionally tackle known racial and ethnic disparities in resource allocation (e.g., kidney transplant allocation) or disparities in care (e.g., prostate cancer screening that historically led to Black men receiving more low-yield biopsies). It is important to note that in such cases the rationale for using race and ethnicity was clearly delineated and did not conflate race and ethnicity with ancestry and/or genetic predisposition. However, when algorithms include race and ethnicity without clear rationale, they may perpetuate the incorrect notion that race is a biologic construct and contribute to disparities. Finally, some algorithms may reduce or perpetuate disparities without containing race and ethnicity as an input. Several modeling studies showed that applying algorithms out of context of original development (e.g., illness severity scores used for crisis standards of care) could perpetuate or exacerbate disparities. On the other hand, algorithms may also reduce disparities by standardizing care and reducing opportunities for implicit bias (e.g., Lung Allocation Score for lung transplantation). Several mitigation strategies have been shown to potentially reduce the contribution of algorithms to racial and ethnic disparities. Results of mitigation efforts are highly context specific, relating to unique combinations of algorithm, clinical condition, population, setting, and outcomes. Important future steps include increasing transparency in algorithm development and implementation, increasing diversity of research and leadership teams, engaging diverse patient and community groups in the development to implementation lifecycle, promoting stakeholder awareness (including patients) of potential algorithmic risk, and investing in further research to assess the real-world effect of algorithms on racial and ethnic disparities before widespread implementation.

Government-owned development banks have often been justified by the need to respond to financial market imperfections that hinder the establishment and growth of promising businesses, and as a result, stifle economic development more generally. However, evidence on the effectiveness of these banks in mitigating financial constraints is still lacking. To fill this gap, this paper analyzes the impact of Bancoldex, Colombia's publicly owned development bank, on access to credit. It uses a unique dataset that contains key characteristics of all loans issued to businesses in Colombia, including the financial intermediary through which the loan was granted and whether the loan was funded with Bancoldex resources. The paper assesses effects on access to credit by comparing Bancoldex loans to loans from other sources and study the impact of receiving credit from Bancoldex on a firm's subsequent credit history. To address concerns about selection bias, it uses a combination of models that control for fixed effects and matching techniques. The findings herein show that credit relationships involving Bancoldex funding are characterized by lower interest rates, larger loans, and loans with longer terms. These characteristics translated into lower average interest rates and larger average loans for firms that used Bancoldex credit. Average loans of Bancoldex' beneficiaries also exhibit longer terms, although this effect can take two years to materialize. Finally, the findings show evidence of a demonstration effect of Bancoldex: beneficiary firms that have access Bancoldex credit are able to significantly expand the number of intermediaries with whom they have credit relationships.