Search

User modeling is a key topic in many applications, mainly social networks and information retrieval systems. To assess the effectiveness of a user modeling approach, its capability to classify personal characteristics (e.g., the gender, age, or consumption grade of the users) is evaluated. Due to the fact that some of the attributes to predict are multiclass (e.g., age usually encompasses multiple ranges), assessing fairness in user modeling becomes a challenge since most of the related metrics work with binary attributes. As a workaround, the original multiclass attributes are usually binarized to meet standard fairness metrics definitions where both the target class and sensitive attribute (such as gender or age) are binary. However, this alters the original conditions, and fairness is evaluated on classes that differ from those used in the classification. In this article, we extend the definitions of four existing fairness metrics (related to disparate impact and disparate mistreatment) from binary to multiclass scenarios, considering different settings where either the target class or the sensitive attribute includes more than two groups. Our work endeavors to bridge the gap between formal definitions and real use cases in bias detection. The results of the experiments, conducted on four real-world datasets by leveraging two state-of-the-art graph neural network-based models for user modeling, show that the proposed generalization of fairness metrics can lead to a more effective and fine-grained comprehension of disadvantaged sensitive groups and, in some cases, to a better analysis of machine learning models originally deemed to be fair. The source code and the preprocessed datasets are available at the following link: . [ABSTRACT FROM AUTHOR]

Postauthorization observational studies play a key role in understanding COVID-19 vaccine effectiveness following the demonstration of efficacy in clinical trials. Although bias due to confounding, selection bias, and misclassification can be mitigated through careful study design, unmeasured confounding is likely to remain in these observational studies. Phase III trials of COVID-19 vaccines have shown that protection from vaccination does not occur immediately, meaning that COVID-19 risk should be similar in recently vaccinated and unvaccinated individuals, in the absence of confounding or other bias. Several studies have used the estimated effectiveness among recently vaccinated individuals as a negative control exposure to detect bias in vaccine effectiveness estimates. In this paper, we introduce a theoretical framework to describe the interpretation of such a bias indicator in test-negative studies, and outline strong assumptions that would allow vaccine effectiveness among recently vaccinated individuals to serve as a negative control exposure.

The sharing of biased information has become an increasingly pervasive issue. This is quite dangerous considering how the exchange of information can influence perceptions, decision-making, and, most importantly, how well we coexist. Accordingly, as our access to information and interactions grow in the wake of the digital age, we must reestablish control for how information is shared and increase accountability for those sharing information. Unfortunately, this is seemingly impossible given the scale of interactions and the complexity of information passed around. Thus, researchers in an experiment by Westmark et al., suggest that gauging people’s ability to detect biased information at the lower group level is where to start to accomplish these initiatives. The present work is an exploratory data analysis of the results from a survey deployed during this experiment, which was used to assess participants’ ability to detect bias correctly. The analysis was designed to provide researchers with different perspectives of the original hypothesis to consider. Although no significant relationships were found, comparisons based on gender and party affiliation displayed interesting information about how well these groups deal with the media information they receive and pass on to others.

Systematic reviews are critical summaries of the exiting literature on a given subject and, when combined with meta-analysis, provides a quantitative synthesis of evidence to direct and inform future research. Such reviews must, however, account for complex sources of between study heterogeneity and possible sources of bias, such as publication bias. This paper presents the methods and results of a research study using a newly developed software tool called ABCal (version 1.0.2) to compute and assess author bias in the literature, providing a quantitative measure for the possible effect of overrepresented authors introducing bias to the overall interpretation of the literature. ABCal includes a new metric referred to as author bias, which is a measure of potential biases per paper when the frequency or proportions of contributions from specific authors are considered. The metric is able to account for a significant portion of the observed heterogeneity between studies included in meta-analyses. A meta-regression between observed effect measures and author bias values revealed that higher levels of author bias were associated with higher effect measures while lower author bias was evident for studies with lower effect measures. Furthermore, the software's capabilities to analyse authorship contributions and produce scientometric plots was able to reveal distinct patterns in both the temporal and geographic distributions of publications, which may relate to any evident publication bias. Thus, ABCal can aid researchers in gaining a deeper understanding of the research landscape and assist in identifying both key contributors and holistic research trends. [ABSTRACT FROM AUTHOR]

In this paper, firstly, we study the reinforcement learning algorithm, which is a related technology of knowledge graph, and use reinforcement learning to structure the application framework of health care knowledge graph and construct the domain dictionary. And based on the BitLab21 dataset, entity relationship annotation is performed on the named entity recognition result set to realize entity relationship extraction of clinical electronic medical records. Then, the HacRED dataset is used to evaluate the deep learning model, and comparison tests and ablation experiments are conducted to verify the effectiveness of reinforcement learning for constructing knowledge graphs, respectively. Finally, a deviation monitoring method and a data feature extraction method for the case dataset are proposed, and a comparative study of different feature extraction methods is conducted to illustrate the advantages of feature fusion in distinguishing abnormal deviation patterns. All three features have relatively good recognition accuracy for normal patterns, which can reach 100%. The error curves perform well, and their training and testing errors converge very quickly to below 0.01 with less than 5 iterations and achieve nearly 100% recognition accuracy. Reinforcement learning-based knowledge mapping of case datasets and feature fusion-based data machine bias detection can improve the efficiency of medical detection.

ABSTRACTThe Microbiology Concept Inventory is an assessment tool derived from the fundamental statements created by the American Society for Microbiology. This two-tier, multiple-choice question inventory requires students to choose the most correct answer for each question and provide a brief justification of their reasoning. Educators can utilize this tool to identify common misconceptions held by students and adjust curriculum to address and prevent the persistence of student misconceptions. Over the course of 5 years, the Microbiology Concept Inventory was annually administered to undergraduate students enrolled in entry-level, mid-level, and senior capstone microbiology courses at a mid-western rural university. Analysis was completed to compare course, year, majors and minors, gender, ethnicity, and cumulative GPA. Results of this study showed a significant difference in Microbiology Concept Inventory scores between students with high cumulative GPAs (3.5–4.0) and students with comparatively lower cumulative GPAs (2.5–2.99, 3.0–3.49). Results between the other demographic categories revealed statistically different scores in favor of white students, but no differences in scores between genders. The results suggest evidence of ethnic bias, but no gender bias as measured by the Microbiology Concept Inventory. Additionally, significant differences in scores across cohorts are indicative of improvements in the curricula due to prior targeted changes. Analysis of concept inventory results can guide curriculum changes for course instructors. Implementation of curriculum changes can enrich students’ academic success.

Bias detection in the computer vision field is a necessary task, to achieve fair models. These biases are usually due to undesirable correlations present in the data and learned by the model. Although explainability can be a way to gain insights into model behavior, reviewing explanations is not straightforward. This work proposes a methodology to analyze the model biases without using explainability. By doing so, we reduce the potential noise arising from explainability methods, and we minimize human noise during the analysis of explanations. The proposed methodology combines images of the original distribution with images of potential context biases and analyzes the effect produced in the model's output. For this work, we first presented and released three new datasets generated by diffusion models. Next, we used the proposed methodology to analyze the context impact on the model's prediction. Finally, we verified the reliability of the proposed methodology and the consistency of its results. We hope this tool will help practitioners to detect and mitigate potential biases, allowing them to obtain more reliable models. [ABSTRACT FROM AUTHOR]

Copyright of Revista de Derecho y Genoma Humano is the property of Dykinson SL and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

With the increase in usage of machine learning models within many different aspects of customer interactions, it has become very clear that bias detection within associated customer interaction datasets has led to a critical focus on issues such as the identification of bias prior to model building, lack of understanding and transparency within models, and ultimately the prevention of biased predictions or classifications. This has never been more important since the introduction of the EU General Data Protection Regulation (GDPR) and the associated rule of “right of explanation”. In this paper, we survey the state of the art for bias detection, avoidance and mitigation within datasets, and the associated methods and tools available. Our purpose is to establish an understanding of how established customer interaction-based use cases can utilise these techniques. The focus is primarily on tackling the bias in unstructured text data as a pre-process prior to the machine learning model training phase. We hope that this research encourages the further establishment of responsible usage of customer interaction datasets to allow the prevention of bias being introduced into machine learning pipelines and to also allow greater awareness of the potential for further research in this area.

Many of the current recommendation systems are considered to be blackboxes that are tuned to optimize some global objective function. However, their error distribution may differ dramatically among different combinations of attributes, and such algorithms may lead to propagating hidden data biases. Identifying potential disparities in an algorithm's functioning is essential for building recommendation systems in a fair and responsible way. In this work, we propose a model-agnostic technique to automatically detect the combinations of user and item attributes correlated with unequal treatment by the recommendation model. We refer to this technique as the Bias Detection Tree. In contrast to the existing works in this field, our method automatically detects disparities related to combinations of attributes without any a priori knowledge about protected attributes, assuming that relevant metadata is available. Our results on five public recommendation datasets show that the proposed technique can identify hidden biases in terms of four kinds of metrics for multiple collaborative filtering models. Moreover, we adapt a minimax model selection technique to control the trade-off between the global and the worst-case optimizations and improve the recommendation model's performance for biased attributes. [ABSTRACT FROM AUTHOR]

The integration of machine learning in applications provides opportunities for increased efficiency in many organisations. However, the deployment of such systems is often hampered by the lack of insight into how their decisions are reached, resulting in concerns about trust and fairness. In this article, we investigate to what extent the addition of explainable AI components to ML applications can contribute to alleviating these issues. As part of this research, explainable AI functionality was developed for an existing ML model used for mortgage fraud detection at a large international financial institution based in The Netherlands A system implementing local explanation techniques was deployed to support the day-to-day work of fraud detection experts working with the model. In addition, a second system implementing global explanation techniques was developed to support the model management processes involving data-scientists, legal experts and compliance officers. A controlled experiment using actual mortgage applications was carried out to measure the effectiveness of these two systems, using both quantitative and qualitative assessment methods. Our results show that the addition of explainable AI functionality results in a statistically significant improvement in the levels of trust and usability by its daily users. The explainable AI system implementing global interpretability was found to considerably increase confidence in the ability to perform the processes focused on compliance and fairness. In particular, bias detection towards demographic groups successfully aided in the identification and removal of bias towards applicants with a migration background.

Transformer models are now increasingly being used in real-world applications. Indiscriminately using these models as automated tools may propagate biases in ways we do not realize. To responsibly direct actions that will combat this problem, it is of crucial importance that we detect and quantify these biases. Robust methods have been developed to measure bias in non-contextualized embeddings. Nevertheless, these methods fail to apply to contextualized embeddings due to their mutable nature. Our study focuses on the detection and measurement of stereotypical biases associated with gender in the embeddings of T5 and mT5. We quantify bias by measuring the gender polarity of T5's word embeddings for various professions. To measure gender polarity, we use a stable gender direction that we detect in the model's embedding space. We also measure gender bias with respect to a specific downstream task and compare Swedish with English, as well as various sizes of the T5 model and its multilingual variant. The insights from our exploration indicate that the use of a stable gender direction, even in a Transformer's mutable embedding space, can be a robust method to measure bias. We show that higher status professions are associated more with the male gender than the female gender. In addition, our method suggests that the Swedish language carries less bias associated with gender than English, and the higher manifestation of gender bias is associated with the use of larger language models. [ABSTRACT FROM AUTHOR]

Observers are deployed on commercial fishing trips to collect representative samples of discard rates. However, fishers may change their fishing habits when an observer is onboard ("observer effect") or observer programs may over- or under-sample portions of the fleet ("deployment effect"). If the extent of these effects are substantial, observer data will not be representative of unobserved trips, potentially biasing the estimation of discards. This sampling bias can impact catch monitoring, stock assessments, and fishery management. The purpose of this study was to examine the power and error rate of several published methods for detecting an observer effect using a simulation of observer and deployment effects at varying sampling ratios (i.e., observer coverage) for several sample statistics. The simplest methods (t-test and F-test for difference of means and variances) provided an accurate although imprecise estimate of the observer effect size, but only when there were no deployment effects. A generalized linear mixed effects model (GLMM) was also not reliable for detecting small bias, but was not confounded by deployment effects and was relatively robust to changing coverage rates. The most complicated tests involved comparing differences in trip characteristics between subsequent trips for observed-unobserved and unobserved-unobserved pairs. These tests were able to detect smaller observer effects and were not confounded by deployment effects, but were unreliable at high coverage rates (>60%), producing both high false positive and false negative rates. Sensitivity tests also showed differing detection accuracy as the distribution of the metric of interest changed. Thus, the optimal test for detecting an observer effect will depend on the metric of interest, the coverage rate, and whether a deployment effect exists. An example from the New England groundfish fishery is provided to illustrate how conflicting results may be explained. Results should always be considered carefully when declaring that an observer effect is or is not occurring because of the sensitivity of the tests. • Simulations show that detecting fisheries observer effects are not straightforward. • Tests can be sensitive to data distributions and observer coverage rates. • More complex tests are needed when deployment effects may also be present. • Accurate detection of observer and deployment effects may affect policy decisions. [ABSTRACT FROM AUTHOR]

Abstract Background Conflicts of interest (COIs), including those arising from interactions with pharmaceutical companies, may lead to bias in medical data. Although medical students are now requesting more education on COIs and bias, they are still not adequately taught during medical school, and few published courses on this topic exist. The objective of our study was therefore to evaluate the feasibility and effectiveness of a blended-learning course for detecting and avoiding bias in medical data, with a special focus on COIs. Methods We developed a blended learning course on bias detection, COIs, and risk communication. It was piloted in the Fall Semester of 2019/2020 using a pre/post-test design. The primary outcome was a gain in bias detection skills, tested by a novel key feature test. Secondary outcomes were (i) skepticism (tested using an attitude questionnaire), (ii) the intention to manage COIs in a professional way so as to avoid bias (tested using a situational judgment test) and (iii) the course evaluation by the students. Results Seventeen students participated in the study. The key feature test showed a significant improvement in bias detection skills at post-testing, with a difference in means of 3.1 points (95%-CI: 1.7–4.4, p-value:

Although of paramount importance to all societies, the fact that media can be biased is a troubling thought to many people. The problem, however, is by no means easy to solve, given its high subjectivity, thereby leading to a number of different approaches by researchers. In this work, we addressed media bias according to a tripartite model whereby news can suffer from a combination of selective coverage of issues (Selection Bias), disproportionate attention given to specific subjects (Coverage Bias), and the favouring of one side in a dispute (Statement Bias). To do so, we approached the problem within an outlier detection framework, defining bias as a noticeable deviation from some mainstream behaviour. Results show that, in following this methodology, one can not only identify bias in specific outlets, but also determine how that bias comes about, how strong it is, and the way it interacts with other dimensions, thereby rendering a more complete picture of the phenomenon under inspection.

Hyperpartisan news is a kind of news riddled with twisted, untruthful, and often extremely one-sided. This kind of news can spread more successfully than the others. One of the obvious traits of hyperpartisan news content is that it can mimic regular news articles. Most are favour fake news detection algorithms, and there is less research conducted for hyperpartisan news. This research aims to perform classification on the hyperpartisan news using ELMo and bias features. ELMo was used to develop a classification model to perform classification on the BuzzFeed Webis News Corpus dataset. The model uses ELMo embedding with bias word score generated from bias lexicon to train a deep learning model using Tensorflow and Keras. We had compared the final result with two proposed baseline models that utilized ELMo from other research. The discussion section further investigated the contribution of ELMo and bias feature in the hyperpartisan task. [ABSTRACT FROM AUTHOR]

Machine learning models are built using training data, which is collected from human experience and is prone to bias. Humans demonstrate a cognitive bias in their thinking and behavior, which is ultimately reflected in the collected data. From Amazon's hiring system, which was built using ten years of human hiring experience, to a judicial system that was trained using human judging practices, these systems all include some element of bias. The best machine learning models are said to mimic humans' cognitive ability, and thus such models are also inclined towards bias. However, detecting and evaluating bias is a very important step for better explainable models. In this work, we aim to explain bias in learning models in relation to humans' cognitive bias and propose a wrapper technique to detect and evaluate bias in machine learning models using an openly accessible dataset from UCI Machine Learning Repository. In the deployed dataset, the potentially biased attributes (PBAs) are gender and race. This study introduces the concept of alternation functions to swap the values of PBAs, and evaluates the impact on prediction using KL divergence. Results demonstrate females and Asians to be associated with low wages, placing some open research questions for the research community to ponder over. [ABSTRACT FROM AUTHOR]

In this work, we introduce and analyze an approach to knowledge transfer from one collection of facts to another without the need for entity or relation matching. The method works for both canonicalized knowledge bases and uncanonicalized or open knowledge bases , i.e., knowledge bases where more than one copy of a real-world entity or relation may exist. The main contribution is a method that can make use of large-scale pre-training on facts, which were collected from unstructured text, to improve predictions on structured data from a specific domain. The introduced method is most impactful on small datasets such as ReVerb20k , where a 6% absolute increase of mean reciprocal rank and 65% relative decrease of mean rank over the previously best method was achieved, despite not relying on large pre-trained models like Bert. To understand the obtained pre-trained models better, we then introduce a novel dataset for the analysis of pre-trained models for Open Knowledge Base Completion, called Doge (Diagnostics of Open knowledge Graph Embeddings). It consists of 6 subsets and is designed to measure multiple properties of a pre-trained model: robustness against synonyms, ability to perform deductive reasoning, presence of gender stereotypes, consistency with reverse relations, and coverage of different areas of general knowledge. Using the introduced dataset, we show that the existing OKBC models lack consistency in presence of synonyms and inverse relations and are unable to perform deductive reasoning. Moreover, their predictions often align with gender stereotypes, which persist even when presented with counterevidence. We additionally investigate the role of pre-trained word embeddings and demonstrate that avoiding biased word embeddings is not a sufficient measure to prevent biased behavior of OKBC models. [ABSTRACT FROM AUTHOR]

Bias in textual data can lead to skewed interpretations and outcomes when the data is used. These biases could perpetuate stereotypes, discrimination, or other forms of unfair treatment. An algorithm trained on biased data may end up making decisions that disproportionately impact a certain group of people. Therefore, it is crucial to detect and remove these biases to ensure the fair and ethical use of data. To this end, we develop a comprehensive and robust framework Nbias that consists of four main layers: data, corpus construction, model development and an evaluation layer. The dataset is constructed by collecting diverse data from various domains, including social media, healthcare, and job hiring portals. As such, we applied a transformer-based token classification model that is able to identify bias words/phrases through a unique named entity BIAS. In the evaluation procedure, we incorporate a blend of quantitative and qualitative measures to gauge the effectiveness of our models. We achieve accuracy improvements ranging from 1% to 8% compared to baselines. We are also able to generate a robust understanding of the model functioning. The proposed approach is applicable to a variety of biases and contributes to the fair and ethical use of textual data. • Nbias comprises data, model development, evaluations. • Nbias classifier outperforms baseline models by 1–8. • Nbias is generalizabile multiple bias dimensions. • Experiments show Nbias identifies various biases. [ABSTRACT FROM AUTHOR]

Background: Conflicts of interest (COIs), including those arising from interactions with pharmaceutical companies, may lead to bias in medical data. Although medical students are now requesting more education on COIs and bias, they are still not adequately taught during medical school, and few published courses on this topic exist. The objective of our study was therefore to evaluate the feasibility and effectiveness of a blended-learning course for detecting and avoiding bias in medical data, with a special focus on COIs. Methods: We developed a blended learning course on bias detection, COIs, and risk communication. It was piloted in the Fall Semester of 2019/2020 using a pre/post-test design. The primary outcome was a gain in bias detection skills, tested by a novel key feature test. Secondary outcomes were (i) skepticism (tested using an attitude questionnaire), (ii) the intention to manage COIs in a professional way so as to avoid bias (tested using a situational judgment test) and (iii) the course evaluation by the students. Results: Seventeen students participated in the study. The key feature test showed a significant improvement in bias detection skills at post-testing, with a difference in means of 3.1 points (95%-CI: 1.7–4.4, p-value: < 0.001; highest possible score: 16 points). The mean score after the course was 6.21 (SD: 2.62). The attitude questionnaire and situational judgment test also showed an improvement in skepticism and intentions to manage COIs, respectively. Students evaluated the course as having been worthwhile (Median: 5, IQR: 0.75, Likert-Scale 1–6, 6 = fully applicable). Conclusions: The blended learning format of the course was feasible and effective. The results suggest a relevant learning gain; however, the low mean score on the key feature test after the course reflects the difficulty of the subject matter. Although a single course has the potential to induce significant short-term improvements in bias detection skills, the complexity of this important subject necessitates its longitudinal integration into medical curricula. This concept should include specific courses such as that presented here as well as an integration of the topic into clinical courses to improve context-related understanding of COIs and medical data bias. [ABSTRACT FROM AUTHOR]

Retail shoplifting is one of the most prevalent forms of theft and has accounted for over one billion GBP in losses for UK retailers in 2018. An automated approach to detecting behaviours associated with shoplifting using surveillance footage could help reduce these losses. Until recently, most state-of-the-art vision-based approaches to this problem have relied heavily on the use of black box deep learning models. While these models have been shown to achieve very high accuracy, this lack of understanding on how decisions are made raises concerns about potential bias in the models. This limits the ability of retailers to implement these solutions, as several high-profile legal cases have recently ruled that evidence taken from these black box methods is inadmissible in court. There is an urgent need to develop models which can achieve high accuracy while providing the necessary transparency. One way to alleviate this problem is through the use of social signal processing to add a layer of understanding in the development of transparent models for this task. To this end, we present a social signal processing model for the problem of shoplifting prediction which has been trained and validated using a novel dataset of manually annotated shoplifting videos. The resulting model provides a high degree of understanding and achieves accuracy comparable with current state of the art black box methods.

Machine learning models are built using training data, which is collected from human experience and is prone to bias. Humans demonstrate a cognitive bias in their thinking and behavior, which is ultimately reflected in the collected data. From Amazon’s hiring system, which was built using ten years of human hiring experience, to a judicial system that was trained using human judging practices, these systems all include some element of bias. The best machine learning models are said to mimic humans’ cognitive ability, and thus such models are also inclined towards bias. However, detecting and evaluating bias is a very important step for better explainable models. In this work, we aim to explain bias in learning models in relation to humans’ cognitive bias and propose a wrapper technique to detect and evaluate bias in machine learning models using an openly accessible dataset from UCI Machine Learning Repository. In the deployed dataset, the potentially biased attributes (PBAs) are gender and race. This study introduces the concept of alternation functions to swap the values of PBAs, and evaluates the impact on prediction using KL divergence. Results demonstrate females and Asians to be associated with low wages, placing some open research questions for the research community to ponder over.

The Microbiology Concept Inventory is an assessment tool derived from the fundamental statements created by the American Society for Microbiology. This two-tier, multiple-choice question inventory requires students to choose the most correct answer for each question and provide a brief justification of their reasoning. Educators can utilize this tool to identify common misconceptions held by students and adjust curriculum to address and prevent the persistence of student misconceptions. Over the course of 5 years, the Microbiology Concept Inventory was annually administered to undergraduate students enrolled in entry-level, mid-level, and senior capstone microbiology courses at a mid-western rural university. Analysis was completed to compare course, year, majors and minors, gender, ethnicity, and cumulative GPA. Results of this study showed a significant difference in Microbiology Concept Inventory scores between students with high cumulative GPAs (3.5-4.0) and students with comparatively lower cumulative GPAs (2.5-2.99, 3.0-3.49). Results between the other demographic categories revealed statistically different scores in favor of white students, but no differences in scores between genders. The results suggest evidence of ethnic bias, but no gender bias as measured by the Microbiology Concept Inventory. Additionally, significant differences in scores across cohorts are indicative of improvements in the curricula due to prior targeted changes. Analysis of concept inventory results can guide curriculum changes for course instructors. Implementation of curriculum changes can enrich students' academic success., Competing Interests: The authors declare no conflict of interest., (Copyright © 2023 Rennpferd et al.)

Research suggests that intergroup disagreement about the prevalence of subtle racial bias in America occurs because Whites are not often exposed to minorities' experiences with this type of discrimination, due to housing, school, workplace, and friendship segregation. Although the objective of social movements (e.g., “ I , too , am Harvard ”) is to illuminate a consistent pattern of bias and to spur social action, it is possible that these movements may exacerbate the derogatory judgments (i.e., as complainers) found in previous research when a single claimant describes experiences with bias. Five studies are the first to draw on the consensus and consistency principles of Kelley's Covariation Model (1973) to investigate how exposure to multiple experiences of subtle bias brought by Black or White claimants affects Whites' perceptions of subsequent discrimination claimants and racial bias prevalence. The results supported the consensus and consistency hypothesis for Black claimants, as increased exposure to Blacks' discrimination experiences mitigated Whites' derogation of Black discrimination claimants as complainers and increased perceptions of the prevalence of anti-Black bias. Conversely, increased exposure to Whites' discrimination experiences supported the derogation hypothesis : exposure exacerbated complainer attributions for those claimants and had no effect on the perceived prevalence of anti-White bias. These results suggest increased exposure may be an effective tool for changing Whites' perceptions of and attitudes toward minorities' subtle bias experiences. We also discuss the contribution of these studies to our understanding of differences between intergroup and intragroup perceptions of discrimination. [ABSTRACT FROM AUTHOR]