Matching Items (10)
Filtering by
- Creators: Arizona State University
- Creators: Liu, Huan
Description
Machine learning models can pick up biases and spurious correlations from training data and projects and amplify these biases during inference, thus posing significant challenges in real-world settings. One approach to mitigating this is a class of methods that can identify filter out bias-inducing samples from the training datasets to force models to avoid being exposed to biases. However, the filtering leads to a considerable wastage of resources as most of the dataset created is discarded as biased. This work deals with avoiding the wastage of resources by identifying and quantifying the biases. I further elaborate on the implications of dataset filtering on robustness (to adversarial attacks) and generalization (to out-of-distribution samples). The findings suggest that while dataset filtering does help to improve OOD(Out-Of-Distribution) generalization, it has a significant negative impact on robustness to adversarial attacks. It also shows that transforming bias-inducing samples into adversarial samples (instead of eliminating them from the dataset) can significantly boost robustness without sacrificing generalization.
ContributorsSachdeva, Bhavdeep Singh (Author) / Baral, Chitta (Thesis advisor) / Liu, Huan (Committee member) / Yang, Yezhou (Committee member) / Arizona State University (Publisher)
Created2021
Description
Street-level bureaucracy (SLB) theory argues that public servants take shortcuts when making decisions about the delivery of public services. These shortcuts can lead SLBs to treat citizens unfairly. Public administration and political science researchers have found some evidence that street-level bureaucrats act in biased ways towards ethnic and racial minorities, citizens of lower socioeconomic status, and religious minorities. I expand on the SLB literature on discrimination by examining whether SLBs discriminate based on the political ideology of citizens. According to the Ideological-Conflict Hypothesis, individuals act in biased ways towards others whose political values conflict with their own. Using the Ideological-Conflict Hypothesis, I test whether SLBs working in local governments discriminate against citizens based on political ideology and whether discrimination is related to type of service delivery (e.g. needs based versus universal). I carry out two audit experiments to test for discrimination. One audit experiment tests for political ideology discrimination in a need-based program among a sample of public housing authorities in the United States (US). The sample is limited to areas where over 60% of citizens voted for the Democratic candidate in the 2020 Presidential Election (n = 274)—and where over 60% voted for the Republican candidate (n = 274). The other audit experiment tests for political ideology discrimination in the delivery of a universal service using a sample municipal parks departments in US cities. The sample is cities with over 25,000 residents where at least 60% of citizens in the county voted for the Democratic candidate in the 2020 Presidential Election (n = 227) and counties where at least 60% of citizens voted for Republican candidate (n = 227). The treatment signals that an email is from a conservative citizen, a liberal citizen, or a citizen with no identifiable political ideology. The results of my dissertation provide some support for the Ideological-Conflict Hypothesis and evidence indicates SLBs discriminate based on political ideology. The results do not find differences in political discrimination for needs-based public service delivery compared to universal public service delivery.
ContributorsOlsen, Jared (Author) / Feeney, Mary K. (Thesis advisor) / Favero, Nathan (Committee member) / Miller, Susan M. (Committee member) / Stritch, Justin (Committee member) / Arizona State University (Publisher)
Created2023
Description
To reveal opinions people may not explicitly portray, researchers have implemented a methodology called the Implicit Associations Test (IAT). While this test saw widespread use after its inception, recent problems have undermined the reliability of the measure. Researchers have begun to address these limitations by evaluating different approaches, such as the Action Dynamics paradigm. Like the IAT, the aim of action dynamics is to assess underlying activation and competition amongst beliefs as they unfold in real-time, while adding a number of more sensitive measures, in addition to those used in an IAT. The trajectories of participants’ computer mouse cursors are tracked as they move from a stimulus statement to a response, providing data of the real-time decisions people are making across a number of variables. For this thesis study, the aim was to use an action dynamics paradigm to explore whether implicit biases exist toward transgender people from a larger cisgender population, even if they explicitly support or oppose others with transgender identities. These potential biases were assessed by evaluating the statements people were asked to confirm or disconfirm. There were also a number of analyses conducted in order to investigate whether such predictors such as
participants’ gender or political ideology predicted differences in responses. Although differences were seen in the reaction time to statements of a certain category, the other trajectory measures showed that participants’ implicit and explicit attitudes toward transgender people were aligned. Implications, limitations, and future directions of this
work are then discussed.
ContributorsHamlett, Mara Carol (Author) / Duran, Nicholas (Thesis advisor) / Mickelson, Kristin (Committee member) / Hall, Deborah (Committee member) / Arizona State University (Publisher)
Created2022
Description
Cultivation theory states that consuming television cultivates a social reality in the real world which aligns with the reality present in television. When the television show CSI was released, researchers studied a form of cultivation stemming from the show titled the "CSI Effect." One of the components of the CSI Effect is the tendency of those who watch CSI to be more likely to overestimate the presence of forensic evidence present in a trial and place more trust in such evidence. In recent years, several true crime documentaries that examined controversial cases have been released. In a similar vein of research conducted on CSI, the current study examines true crime documentaries and their possible impacts on viewers’ judgments and beliefs about the criminal justice system. In the current study, participants were provided with a mock case and asked about their perceptions of the case along with their viewership habits. While overall true crime documentary viewership did not influence judgments of evidence manipulation or perceptions of police, findings point to viewership of the targeted documentaries being associated with feelings of mistrust towards the criminal justice system overall, while the lesser-viewed documentaries correlated with judgments of strength and responsibility of the defendant in the case. One possible explanation is that individual characteristics may serve as the driving factor in how individuals choose what to watch when the popularity of the show is not as well-known.
ContributorsDoughty, Kathryn A (Author) / Schweitzer, Nicholas J. (Thesis advisor) / Neal, Tess (Committee member) / Salerno, Jessica (Committee member) / Arizona State University (Publisher)
Created2018
Description
Exabytes of data are created online every day. This deluge of data is no more apparent than it is on social media. Naturally, finding ways to leverage this unprecedented source of human information is an active area of research. Social media platforms have become laboratories for conducting experiments about people at scales thought unimaginable only a few years ago.
Researchers and practitioners use social media to extract actionable patterns such as where aid should be distributed in a crisis. However, the validity of these patterns relies on having a representative dataset. As this dissertation shows, the data collected from social media is seldom representative of the activity of the site itself, and less so of human activity. This means that the results of many studies are limited by the quality of data they collect.
The finding that social media data is biased inspires the main challenge addressed by this thesis. I introduce three sets of methodologies to correct for bias. First, I design methods to deal with data collection bias. I offer a methodology which can find bias within a social media dataset. This methodology works by comparing the collected data with other sources to find bias in a stream. The dissertation also outlines a data collection strategy which minimizes the amount of bias that will appear in a given dataset. It introduces a crawling strategy which mitigates the amount of bias in the resulting dataset. Second, I introduce a methodology to identify bots and shills within a social media dataset. This directly addresses the concern that the users of a social media site are not representative. Applying these methodologies allows the population under study on a social media site to better match that of the real world. Finally, the dissertation discusses perceptual biases, explains how they affect analysis, and introduces computational approaches to mitigate them.
The results of the dissertation allow for the discovery and removal of different levels of bias within a social media dataset. This has important implications for social media mining, namely that the behavioral patterns and insights extracted from social media will be more representative of the populations under study.
Researchers and practitioners use social media to extract actionable patterns such as where aid should be distributed in a crisis. However, the validity of these patterns relies on having a representative dataset. As this dissertation shows, the data collected from social media is seldom representative of the activity of the site itself, and less so of human activity. This means that the results of many studies are limited by the quality of data they collect.
The finding that social media data is biased inspires the main challenge addressed by this thesis. I introduce three sets of methodologies to correct for bias. First, I design methods to deal with data collection bias. I offer a methodology which can find bias within a social media dataset. This methodology works by comparing the collected data with other sources to find bias in a stream. The dissertation also outlines a data collection strategy which minimizes the amount of bias that will appear in a given dataset. It introduces a crawling strategy which mitigates the amount of bias in the resulting dataset. Second, I introduce a methodology to identify bots and shills within a social media dataset. This directly addresses the concern that the users of a social media site are not representative. Applying these methodologies allows the population under study on a social media site to better match that of the real world. Finally, the dissertation discusses perceptual biases, explains how they affect analysis, and introduces computational approaches to mitigate them.
The results of the dissertation allow for the discovery and removal of different levels of bias within a social media dataset. This has important implications for social media mining, namely that the behavioral patterns and insights extracted from social media will be more representative of the populations under study.
ContributorsMorstatter, Fred (Author) / Liu, Huan (Thesis advisor) / Kambhampati, Subbarao (Committee member) / Maciejewski, Ross (Committee member) / Carley, Kathleen M. (Committee member) / Arizona State University (Publisher)
Created2017
Description
Juror impartiality is necessary for a fair and just legal system, but is true juror impartiality
realistic? The current study investigated the role of implicit and explicit social-cognitive biases in jurors’ conceptualizations of insanity, and the influence of those biases in juror verdict decisions. It was hypothesized that by analyzing the role of implicit and explicit biases in insanity defense cases, jurors’ attitudes towards those with mental illnesses and attitudes towards the insanity defense would influence jurors’ final verdict decisions. Two hundred and two participants completed an online survey which included a trial vignette incorporating an insanity defense (adapted from Maeder et al., 2016), the Insanity Defense Attitude Scale (Skeem, Louden, & Evans, 2004), Community Attitudes Towards the Mentally Ill Scale (Taylor & Dear, 1981), and an Implicit Association Test (Greenwald et al., 1998). While implicit associations concerning mental illness and dangerousness were significantly related to mock jurors’ verdicts, they no longer were when explicit insanity defense attitudes were added to a more complex model including all measured attitudes and biases. Insanity defense attitudes were significantly related to jurors’ verdicts over and above attitudes about the mentally ill and implicit biases concerning the mentally ill. The potentially biasing impact of jurors’ insanity defense attitudes and the impact of implicit associations about the mentally ill in legal judgments are discussed.
realistic? The current study investigated the role of implicit and explicit social-cognitive biases in jurors’ conceptualizations of insanity, and the influence of those biases in juror verdict decisions. It was hypothesized that by analyzing the role of implicit and explicit biases in insanity defense cases, jurors’ attitudes towards those with mental illnesses and attitudes towards the insanity defense would influence jurors’ final verdict decisions. Two hundred and two participants completed an online survey which included a trial vignette incorporating an insanity defense (adapted from Maeder et al., 2016), the Insanity Defense Attitude Scale (Skeem, Louden, & Evans, 2004), Community Attitudes Towards the Mentally Ill Scale (Taylor & Dear, 1981), and an Implicit Association Test (Greenwald et al., 1998). While implicit associations concerning mental illness and dangerousness were significantly related to mock jurors’ verdicts, they no longer were when explicit insanity defense attitudes were added to a more complex model including all measured attitudes and biases. Insanity defense attitudes were significantly related to jurors’ verdicts over and above attitudes about the mentally ill and implicit biases concerning the mentally ill. The potentially biasing impact of jurors’ insanity defense attitudes and the impact of implicit associations about the mentally ill in legal judgments are discussed.
ContributorsHamza, Cassandra (Author) / Neal, Tess M.S. (Thesis advisor) / Schweitzer, Nicholas (Committee member) / Hall, Deborah (Committee member) / Arizona State University (Publisher)
Created2018
Description
Generative Adversarial Networks are designed, in theory, to replicate the distribution of the data they are trained on. With real-world limitations, such as finite network capacity and training set size, they inevitably suffer a yet unavoidable technical failure: mode collapse. GAN-generated data is not nearly as diverse as the real-world data the network is trained on; this work shows that this effect is especially drastic when the training data is highly non-uniform. Specifically, GANs learn to exacerbate the social biases which exist in the training set along sensitive axes such as gender and race. In an age where many datasets are curated from web and social media data (which are almost never balanced), this has dangerous implications for downstream tasks using GAN-generated synthetic data, such as data augmentation for classification. This thesis presents an empirical demonstration of this phenomenon and illustrates its real-world ramifications. It starts by showing that when asked to sample images from an illustrative dataset of engineering faculty headshots from 47 U.S. universities, unfortunately skewed toward white males, a DCGAN’s generator “imagines” faces with light skin colors and masculine features. In addition, this work verifies that the generated distribution diverges more from the real-world distribution when the training data is non-uniform than when it is uniform. This work also shows that a conditional variant of GAN is not immune to exacerbating sensitive social biases. Finally, this work contributes a preliminary case study on Snapchat’s explosively popular GAN-enabled “My Twin” selfie lens, which consistently lightens the skin tone for women of color in an attempt to make faces more feminine. The results and discussion of the study are meant to caution machine learning practitioners who may unsuspectingly increase the biases in their applications.
ContributorsJain, Niharika (Author) / Kambhampati, Subbarao (Thesis advisor) / Liu, Huan (Committee member) / Manikonda, Lydia (Committee member) / Arizona State University (Publisher)
Created2020
Description
Biases have been studied in many legal contexts, including sexual assault cases. Sexual assault cases are complex because there are many stages that biases can come into play and have lasting effects on the rest of the case proceedings. One aspect that has not been widely explored is how people perceive institutions’ liability in sexual assault cases based on an obligation to create non-discriminating environments for members and employees according to laws like Title VII and Title IX. The current project focused on how and why cognitive biases affect laypeople’s judgment. Specifically, laypeople’s ability to discern the strength of evidence in civil sexual assault cases against institutions. This was addressed in a series of two studies, with samples collected from Prolific Academic (n = 90) and Arizona State University students (n = 188) for Study 1 (N = 278), and Prolific Academic in Study 2 (N = 449). Both studies used Latin-square design methods, with within and between subject elements, looking at how confirmation bias influenced decisions about whether an institution demonstrated negligence, and thus liability, in the way they responded to sexual assault allegations within their institution. Results from these studies suggest that jurors are overall accurately able to differentiate between weak and strong cases. However, consistent with previous literature, jurors may be susceptible to confirmation bias from outside information (e.g., news stories) and negatively influenced by their personal attitudes (e.g., rape myth acceptance). Given the increased attention of the Me Too movement, these results provide an initial insight into how individuals may be judging these types of cases against institutions.
ContributorsMcCowan, Kristen (Author) / Neal, Tess M.S. (Thesis advisor) / Salerno, Jessica M (Committee member) / Davis, Kelly C (Committee member) / Arizona State University (Publisher)
Created2020
Description
Cognitive heuristics, or mental shortcuts, sometimes give rise to biases that can influence decision making. These biases may be particularly impactful in a legal context where decision making has lifelong consequences. One such legal decision falls upon social workers who are often tasked with providing custodial recommendations in child custody cases. Across a series of 2 studies, I explored the role of confirmation bias in social worker decision making, the potential value of blinding to reduce bias, as well as social workers’ perceptions of their own biases. Social workers were given detailed case materials describing a custody case between the state and a father. Participants were randomly assigned to read a previous examiner’s positive evaluation of a father, a negative evaluation of the father, or were blinded to a previous examiners rating. Social workers engaged in confirmation bias, such that those who read a positive evaluation of the father viewed him more positively than participants who read a negative evaluation of the father, despite the fact that all of the actual case evidence remained constant. Blinding did not appear to mitigate the bias. In study 2, social workers viewed themselves as less biased than their peers and less biased than other experts in a different field – signifying the presence of a bias blindspot. Together, my findings suggest the need to further explore how bias might affect judgments and also how to mitigate biases, such as making experts aware of their potential for bias.
ContributorsDenne, Emily (Author) / Neal, Tess M.S. (Thesis advisor) / Stolzenberg, Stacia N. (Committee member) / Fabricius, William (Committee member) / Arizona State University (Publisher)
Created2021
Description
ABSTRACT
Psychological assessments contain important diagnostic information and are central to therapeutic service delivery. Therapists' personal biases, invalid cognitive schemas, and emotional reactions can be expressed in the language of the assessments they compose, causing clients to be cast in an unfavorable light. Logically, the opinions of subsequent therapists may then be influenced by reading these assessments, resulting in negative attitudes toward clients, inaccurate diagnoses, adverse experiences for clients, and poor therapeutic outcomes. However, little current research exists that addresses this issue. This study analyzed the degree to which strength-based, deficit-based, and neutral language used in psychological assessments influenced the opinions of counselor trainees (N= 116). It was hypothesized that participants assigned to each type of assessment would describe the client using adjectives that closely conformed to the language used in the assessment they received. The hypothesis was confirmed (p = .000), indicating significant mean differences between all three groups. Limitations and implications of the study were identified and suggestions for further research were discussed.
Psychological assessments contain important diagnostic information and are central to therapeutic service delivery. Therapists' personal biases, invalid cognitive schemas, and emotional reactions can be expressed in the language of the assessments they compose, causing clients to be cast in an unfavorable light. Logically, the opinions of subsequent therapists may then be influenced by reading these assessments, resulting in negative attitudes toward clients, inaccurate diagnoses, adverse experiences for clients, and poor therapeutic outcomes. However, little current research exists that addresses this issue. This study analyzed the degree to which strength-based, deficit-based, and neutral language used in psychological assessments influenced the opinions of counselor trainees (N= 116). It was hypothesized that participants assigned to each type of assessment would describe the client using adjectives that closely conformed to the language used in the assessment they received. The hypothesis was confirmed (p = .000), indicating significant mean differences between all three groups. Limitations and implications of the study were identified and suggestions for further research were discussed.
ContributorsScott, Angela N (Author) / Kinnier, Richard (Thesis advisor) / Homer, Judith (Committee member) / Kurpius, Sharon (Committee member) / Arizona State University (Publisher)
Created2015