Bias refers to an inclination towards or against a particular individual, group, or sub-groups. AI models may inherit biases from training data or introduce new forms of bias.
More in Detail#
The success of Machine Learning (ML) systems in visual recognition, online advertising, and recommendation systems have inspired its use in applications such as employee hiring, legal systems, social systems, and voice interfaces such as Alexa, Siri, and the like. Along with the proliferation of these domains, a significant concern regarding the trustworthiness of decisions has risen due to various biases (or systematic errors) which may produce skewed results in the automated decision making. The word ‘bias’ has an established normative explanation in legal language, where it refers to ‘judgement based on preconceived notions or prejudices, as opposed to the impartial evaluation of facts’ . In a more generalized version, bias refers to an inclination towards or against a particular individual, group, or sub-groups. The real world is often described as biased in this sense, and since machine learning techniques simply imitate observations of the world, it should come as no surprise that the resulting systems also capture the same bias .
Correctional Offender Management Profiling for Alternative Sanctions (COMPAS)  software for recidivism prediction, used by the U.S department courts to decide whether to release a person or keep them in prison, has discontinued its use after a careful investigation conducted by the U.S officers as they concluded that the software is biased against African-Americans. Also, the use of predictive policing  software has been ceased due to the presence of racial biases. Amazon’s employment hiring  application realized that it is biased against women. Content personalization and ad ranking systems have been accused of filter bubbles and racial and gender profiling. Bidirectional Encoder Representations from Transformers (BERT), has shown signs of gender biases in google search as it is observed that the gender-neutral terms (such as receptionist, doctor, nurse etc.) acquire stereotype and bias due to the context in which they are present in the corpus . In image domain, a latest gender classification report from the National Institute for Standards and Technology (NIST) pointed out that image classification algorithms performed worse for female-labeled faces than male-labeled faces , exhibit gender biases . A bias can exist in different forms and shapes based on the domain and context of application. The main reasons for the origin of these biases are manifold. An outline of bias-inducing stages in the ML pipeline is detailed in . Based on this study, bias definitions can be induced in data, algorithms, and user interaction feedback loops.
ML systems are primarly based on data-driven approaches; therefore, the outcome of ML-based decision-making processes depends on the input data and the interpretation of that data. This decision-making process involves numerous data analyses, such as uncovering patterns in the data, finding correlations and trends, missing data imputations, and data pre-processing. The performance of ML models depends on the data used to train these models and the analysis performed on the training data. It is noted that the primary source of biases is from the data and its processing- involves what data was used for training, how it was collected, and how the data was generated and pre-processed. A general definition of dataset bias is that the data is not representative of the population of study . Nevertheless, in a broader sense, it also occurs when the data does not contain features for specific applications we are interested in. Additionally, human interactions with the data produce bias against a specific group or individual . Various forms of dataset biases have been identified in ML systems. Sample/selection biases emerge due to the non-random sampling of groups and sub-groups. Exclusion bias arises at the data pre-processing stage when valuable data are omitted thought to be unnecessary. When the data used for training a model is different from the data collected from the real world, for example, the training data is collected using a fixed camera in image training, but the production data is collected using different cameras, a measurement bias can be occurred. Recall bias is a kind of measurement bias, and it occurs when similar data are inconsistently labeled. Observer bias occurs when we observe data based on what we want to see or expect to see. Association biases are resultant of the spurious correlations between features in the data.
Furthermore, algorithmic biases are systematic errors in computer systems or models that cause certain privileges in outcomes concerning a particular group or a person. These biases can emerge in various ways. Foremost among these are the design of the algorithm or the way it uses the datasets to be coded, collected, selected, and processed. Algorithmic errors may lead to biased outcomes even though the data used for training are unbiased. A clear example is pre-existing bias, arising as the result of underlying social and institutional ideologies . Another algorithmic bias is caused by technical biases manifested due the technical limitations of code, its computational resources, its design, and the constraints on the system. Technical biases are more frequent when we rely more on the algorithm in other domains or unanticipated contexts. Moreover, and related to the algorithm internals, correlation biases materialise when algorithms assume conclusions from the correlations in data attributes without knowing the specific purpose of those attributes. Finally, another known bias – in between the identified ones – are feedback loop biases. These arise when there is a recursion error in the mechanism in which information is processed into the data-model-experience pipeline.
The skewed outcomes from the biased data or (and) biased algorithms affect user decisions which may result in a more biased data for future ML systems. For example, consider a search engine which ranks queries. The end users interact mostly with the top ranked results, rather than going down the list, that can affect popularity and user interest of the upcoming decisions, due to the biased interactions.
As a long term vision to create responsible ML systems, identifying and mitigating biases throughout the ML development life cycle should be given paramount importance. In a broader sense, the different ways the bias could be mitigated are:
Identify and define potential sources
Set up guidelines and rules for data collection as well as a model use
Define accurate representative data for training
Properly document and share how the entire data collection process has been done
Incorporate ways to measure and mitigate biases as part of the standard evaluation procedure.
Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. A survey on bias and fairness in machine learning. ACM Comput. Surv., 54(6):115:1–115:35, 2021.
A Campolo, M Sanfilippo, M Whittaker, and K Crawford. Ai now 2017 symposium and workshop. AI Now Institute at New York University, 2018.
Thomas Hellström, Virginia Dignum, and Suna Bensch. Bias in machine learning–what is it good for? arXiv preprint arXiv:2004.00686, 2020. URL: https://arxiv.org/abs/2004.00686.
Tim Brennan and William Dieterich. Correctional offender management profiles for alternative sanctions (compas). Handbook of Recidivism Risk/Needs Assessment Tools (2018), 2018.
Kiana Alikhademi, Emma Drobina, Diandra Prioleau, Brianna Richardson, Duncan Purves, and Juan E Gilbert. A review of predictive policing from the perspective of fairness. Artificial Intelligence and Law, pages 1–17, 2021.
Akhil Alfons Kodiyan. An overview of ethical issues in using ai systems in hiring with a case study of amazon’s ai based hiring tool. Researchgate Preprint, 2019.
Moin Nadeem, Anna Bethke, and Siva Reddy. Stereoset: measuring stereotypical bias in pretrained language models. In ACL/IJCNLP (1), 5356–5371. Association for Computational Linguistics, 2021.
Patrick J Grother, Patrick J Grother, and Mei Ngan. Face recognition vendor test (frvt). US Department of Commerce, National Institute of Standards and Technology, 2014.
Tatiana Tommasi, Novi Patricia, Barbara Caputo, and Tinne Tuytelaars. A deeper look at dataset bias. In Domain adaptation in computer vision applications, pages 37–55. Springer, 2017.
Roel Dobbe, Sarah Dean, Thomas Krendl Gilbert, and Nitin Kohli. A broader view on bias in automated decision-making: reflecting on epistemology and dynamics. CoRR, 2018. arXiv:1807.00553.
This entry was written by Resmi Ramachandranpillai, Fredrik Heintz, Miguel Couceiro, and Gabriel Gonzalez-Castañé.