2.1 Definition of fairness and bias in standards and regulations

Standardization bodies such as ISO/IEC and international organizations such as OECD and European Commission provide documents and guidelines on the trustworthiness of artificial intelligence. ISO/IEC TR 24028:2020 ^[1] includes transparency, controllability, robustness, recoverability, fairness, stability, personal data protection, and security as considerations necessary for ensuring trustworthiness, and definitions and related content regarding fairness can be found in ISO/IEC TR 24028:2020, ISO/IEC TR 24027 ^[2], and ISO/IEC 23894:2023 ^[3].

In ^[1], bias is described as one of the factors that can undermine the trustworthiness of artificial intelligence systems, defined as ``Favoritism towards some things, people or groups over others.'' <sup>[2]</sup> explains the causes of unfairness as follows: ``Bias in objective functions, imbalanced data sets, and human biases in training data and in providing feedback to systems, bias issue in the product concept, the problem formulation or choices about when and where to deploy AI systems.'' ^[3] defines bias in AI systems as ``a systematic difference in treatment of certain objects, people, or groups in comparison to others'', and describes fairness as ``a concept that is distinct from, but related to bias.'' When referring to the fairness of artificial intelligence systems, OECD's ``Recommendation of the Council on Artificial Intelligence'' <sup>[4]</sup> and UNESCO's ``Recommendation on the ethics of artificial intelligence ^[5]'' emphasize the need to ensure fairness of the system by striving to prevent discriminatory biased practices and outcomes throughout the entire life cycle.

The US National Institute of Standards and Technology (NIST) explains fairness as one of the fundamental principles related to the risks of artificial intelligence as follows: ``Fairness in AI includes concerns for equality and equity by addressing issues such as harmful bias and discrimination. Standards of fairness can be complex and difficult to define because perceptions of fairness differ among cultures and may shift depending on application <sup>[6]</sup>.'' The Alan Turing Institute in the UK has provided guidelines for applying AI ethics and safety principles to the design and implementation of algorithmic systems in the public sector in their publication ``Understanding artificial intelligence ethics and safety.'' Here, fairness is described as a principle for responsible design and use of AI technology, stating that ``All AI systems that process social or demographic data pertaining to features of human subjects must be designed to meet a minimum threshold of discriminatory non-harm ^[7].''

2.2 Fairness and bias in academia

2.2.1 Bias in artificial intelligent systems

Most artificial intelligence systems and algorithms are based on data and operate by training on this data. Therefore, data is closely related to the operation of artificial intelligence systems. If there is bias in the data used for training, the algorithm learned from that data can generate biased results or amplify the bias in the data. Additionally, even when the data is unbiased, a biased behavior can be exhibited due to specific algorithmic design choices. The results of biased algorithms can affect user decisions and ultimately lead to a vicious cycle of generating biased data for future algorithm training. For example, with web search engines that place certain results at the top of the list, users tend to interact more with the top results and pay less attention to the results at the bottom of the list. User-item interactions collected by the web search engine are then used to make decisions about how to display information based on popularity and user interest. This can create bias towards the top results not because of the characteristics of the results themselves, but due to biased interactions and list placement. The types of bias in ^[17] are divided into three categories: data, algorithm, and user interaction, defined as interacting loops. Fig. 1. illustrates the cause-and-effect relationships among these factors and categorizes them accordingly. Table Table 2 summarizes the types and causes of biases in a table. For more detailed information, refer to ^[17].

Fig. 1. Inherent bias in the feedback loop among data, algorithm, and users.

Table 2. Summary of three types and causes of biases.

Categories	Bias	Causes of occurrence
Data to Algorithm	Measurement	Occurs in the method of selecting, utilizing, and measuring features
	Omitted variable Bias	When one or more important variables are excluded from the model
	Representation Bias	Occurs in the method of sampling from a population
	Aggregation bias	when drawing incorrect conclusions about individuals by observing the entire population
	Simpson's Paradox	A type of aggregate bias that occurs in the analysis of heterogeneous data
	Sampling Bias	Similar to selection bias, it occurs due to non-random sampling of subgroups
	Longitudinal Data Fallacy	In cross-sectional analysis, combining diverse groups at a single point in time can lead to bias, as heterogeneous groups can lead to different conclusions
	Linking Bias	when the social connection properties obtained from user connections, activities, or interactions differ from actual user behavior.
Algorithm to User	Algorithmic Bias	When bias is purely added by the algorithm and not present in the input data
	User Interaction Bias	biased behavior or interactions based on user interfaces and user choices
	Presentation Bias	Bias due to the way information is represented. For example, on the web, users can only click on the visible content, so clicks occur on the displayed content and not on any other content
	Ranking Bias	Top-ranking results are thought to be the most relevant and important, so they induce more clicks. This bias affects search engines and crowdsourcing applications.
	Popularity Bias	More popular items tend to be displayed more frequently. Popularity measures can be manipulated. This type of bias can be seen in recommendation systems, search engines, or other systems where popular items are presented more to the public.
	Emergent Bias	This refers to bias that emerges as a result of interaction with actual users, and typically occurs after a certain period of time has elapsed following the completion of the design, due to changes in population/cultural values or social knowledge.
	Evaluation Bias	Bias that occurs during model evaluation, such as the use of imbalanced benchmarks.
User to Data	Historical Bias	an existing bias that is a socio-technological issue, where search results for "female CEO" are influenced by the fact that only 5% of Fortune 500 CEOs are women.
	Population Bias	when statistics, population distribution, representative characteristics, or user characteristics differ from the platform's user distribution.
	Self-Selection Bias	when research subjects are self-selected, such as passionate supporters being more likely to participate in a poll.
	Social Bias	when others' behavior influences our judgment, such as when we want to rate or review an item poorly but are influenced by others' high ratings.
	Behavioral Bias	when user behavior differs depending on the platform, context, or dataset.
	Temporal Bias	Temporal bias arises from differences in population and behavior over time, such as when people start using hashtags to attract attention to a topic and then continue discussing the event without using hashtags.
	Content Protection Bias	structural, lexical, semantic, and syntactical differences in user-generated content, such as differences in language use based on gender, age, country, and population.

3.1 Pre-processing

The creation of all datasets is the result of various design decisions made by the data creator (curator). These decisions have an impact on the fairness of the dataset and ultimately affect the results of the algorithm. Best practices have been proposed to mitigate data bias, such as creating a datasheet containing information on the dataset's creation method, features, motivations, and skewed distribution of the data ^[18,19]. Labeling, such as nutrition labels on food, has also been suggested for better data classification ^[20]. Studies have been conducted to propose methods for testing Simpson's paradox ^[21] or for automatically detecting bias ^[22,23]. Causal models and causal graphs are also used for tasks that involve detecting direct discrimination in the data ^[24] and preventing techniques to modify the data ^[25]. Furthermore, research has been proposed to prevent discrimination in data mining ^[26], including messaging ^[27], preferential sampling ^[28,29], and disparate impact removal ^[30], which are other preprocessing approaches aimed at removing bias from data. ^[31] suggests a fair data envelopment analysis method, incorporating an additional constraint to mitigate disparate impact, thereby addressing the scoring parity between privileged and unprivileged groups. ^[32] formulates fairness as an optimization problem, obscuring information about protected group members to find a good representation for group and individual fairness during data encoding. ^[33] presents optimized preprocessing, which uses a probabilistic transformation to modify data features and labels with fairness and accuracy constraints for group fairness, individual distortion, and data fidelity. ^[34] performs a theoretical study on the trade-off relationship between accuracy and discrimination and proposes four methods using non-discriminatory constraints in classifier design. ^[35] applies a linear transformation on the non the non-sensitive feature columns to eliminate their correlation with the sensitive feature columns while preserving the maximum amount of information possible.

3.2 In-Processing

Classification is a common task in machine learning used in various domains where human contact is not feasible, and techniques have been proposed to minimize cost while meeting certain definitions of fairness to achieve fair classification ^[36]. These techniques include adding constraints for fairness to the cost function through regularization ^[37,38,39], optimizing the estimation probability of sensitive and non-sensitive features to be equal ^[40], modifying the Naive Bayes classifier ^[41], maximizing the average accuracy of each group ^[42], mitigating dependency on sensitive features using Wasserstein distance measure ^[43], using Preferential Sampling (PS) for discrimination-free training data ^[44], and incorporating attention mechanism for interpretable and bias-mitigated classification ^[45]. The ART classifier is used as a classifier to satisfy fairness such as demographic parity and equalized odds ^[46]. ^[47] is a meta-algorithm that modifies the classification threshold of the base classifier to maximize a fairness metric while maintaining a certain level of accuracy. Adversarial debiasing ^[48] trains a classifier to improve prediction accuracy while simultaneously minimizing the adversary's ability to infer the protected attribute from the predictions. ^[49] proposes a group fairness classifier aiming to avoid the susceptibility of intentional or inadvertent ``fairness gerrymandering'' that can occur with traditional statistical definitions of fairness that fix a small collection of pre-defined groups. Exponentiated Gradient Reduction ^[50] reduce fair classification to a sequence of cost-sensitive classification problems, resulting in a randomized classifier with the lowest empirical error while adhering to fair classification constraints.

In community detection, users with low connectivity are often ignored, which can lead to improper classification into communities. However, CLAN ^[51] first identifies important communities and then uses a method of classifying a small number of users into important communities, thus improving the community classification performance of these minority users. In addition, methods for community detection using graph embedding ^[52] or clustering ^[53,54] have also been proposed.

In regression problems, ^[55] proposed to define the loss (cost) function by dividing individuals, groups, and individuals/groups rather than using regression methods for the entire dataset in order to minimize it. Methods for modeling demographic and group-specific fairness as loss functions ^[56], and using decision trees that can be applied to classification and regression problems have also been proposed ^[57]. ^[58] presented an algorithm-independent lower bound by characterizing the trade-off between the inherent fairness and accuracy in regression problems. ^[59] proposed a two-step approach where an unconstrained regression is first estimated from labeled data and then recalibrated with unlabeled data. ^[60] examines the problem of achieving demographic parity in real-valued function learning by establishing a connection between fair regression and optimal transport theory, and proposes a post-processing algorithm to achieve fairness with established guarantees.

Fair PCA ^[61] was proposed to address the issue of amplifying the reconstruction error of one group in original PCA when multiple groups of equal size exist in the data, and it aims to achieve similar representation across different groups. To avoid the problem of inadequate representation of group data other than the lowest principal component during dimensionality reduction of the entire dataset, this method defines a loss function for each group and only performs PCA for the group with the highest loss function. ^[62] formulated the maximum mean discrepancy (MMD)-based fair PCA as a constrained optimization over the Stiefel manifold and solved it using Riemannian Exact Penalty Method via Smoothing.

In addition to classification and regression, studies aimed at mitigating bias have been proposed in the literature. Prejudice remover ^[63] modifies the learning objective by adding a regularization term that is sensitive to discrimination. ^[64] exploits adversarial learning for mitigating unwanted biases in work embeddings.

Causal models are used to eliminate unwanted dependencies on sensitive attributes such as gender or race by using causal graphs to confirm causal relationships between variables ^[65,66]. In ^[67], a method is proposed for detecting direct/indirect discrimination from historical data using causal graphs and removing discrimination effects before predictive analysis. First, path-specific effects (direct/indirect) are defined as below and discrimination thresholds for protected groups are set. For example, the UK law on gender discrimination in 1975 allowed a 5% difference. Similarly, thresholds for indirect discrimination for Redlining properties can be set. From this, direct and indirect discrimination for all paths in the causal graph are determined, and a objective function is created and constraints for mitigating discrimination are set to modify the causal graph through quadratic programming.

There is research on autoencoders that learn fair representations and avoid interference of sensitive attributes ^[68]. In this research, the goal is to obtain fair representations by removing information about sensitive variables. A maximum mean discrepancy regularizer is used to ensure that the posterior distribution of the latent variables is the same. In ^[69], the optimization objective function is designed to include the mutual information between the encoding and sensitive variables, which helps to remove biases in the representation. Debiased variational autoencoder (DB-VAE) is used to reduce the accuracy gap by training the model to learn sensitive attribute variables such as skin color and gender ^[70]. This method modifies the traditional variational autoencoder (VAE) by incorporating the learning of classes corresponding to the samples and representing the frequency of occurrence of specific classes as a histogram. The sample occurrence probability is defined to be inversely proportional to the frequency, ensuring that during training, the samples do not converge to a single class. Adaptive re-sampling is employed to prevent the network from biasing towards a particular class as the learning progresses.

Fig. 2. Debiasing variational autoencoder (DB-VAE) employs resampling during training, where the sample selection is inversely proportional to its frequency of appearance, to mitigate data bias ^[70].

To mitigate biases from learned models with fixed conceptual associations, adversarial debiasing methods were proposed. The goal here is to enhance prediction accuracy while simultaneously reducing the predictions made by an adversary regarding protected and sensitive variables ^[71]. Instead of removing dependency on discriminatory data, FairGAN ^[72] proposed a method to generate data that is similar to real data but without discrimination. The architecture of FairGAN consists of one generator and two discriminators. The generator generates fake data from protective attributes and noise, while the discriminator plays a role in distinguishing real data from fake data. The third component is for the fairness constraint, determining whether generated (fake) data is created from a protected attribute group.

In the research of word embedding which represents a word to a vector for natural language processing, ^[73] proposed the bias relationships such as ``man to computer programmer'' and ``woman to homemaker.'' This method defined the gender-specific and gender-neutral words, and re-embedded the latter so that gender relationship in the pair is removed. Recently, ^[74] proposed a method to analyze and mitigate the biases of contextualized vectors.

Recent advancements in reinforcement learning have enabled automated online decision-making in communication network design and optimization problems. However, most reinforcement learning approaches prioritize maximizing agent rewards without considering fairness. Recent research proposes introducing a fairness term in actor-critic reinforcement learning, adjusting rewards multiplicatively ^[75]. In this study, weighted multiplicative-adjusted rewards are given to the Actor instead of traditional rewards, addressing fairness concerns in network utility optimization.

3.3 Post-Processing

Post-processing techniques represent a subset of methods for mitigating fairness issues in machine learning. These methods are applied subsequent to the model making predictions, with the goal of adjusting the model's outputs to ensure fairness through various means.

The ``Equalized Odds Postprocessing'' technique modifies predicted probabilities for distinct groups to meet the equalized odds criterion <sup>[76,77]</sup>. Its objective is to guarantee that the odds of accurate classification are equivalent across different demographic or sensitive attribute groups. Expanding on the equalized odds concept, the ``Calibrated Equality Odds'' method involves calibrating the model's predicted probabilities and subsequently applying equalized odds postprocessing ^[77]. The aim is to concurrently achieve calibration and fairness. The ``Reject Option Classification'' approach <sup>[78]</sup> introduces a ``reject'' option, where the model refrains from making predictions in situations of high uncertainty. This strategy seeks to alleviate biases by yielding more favorable outcomes to underprivileged groups and less favorable outcomes to privileged groups within a confidence band around the decision boundary ^[79]. In a binary classification context, ^[79] proposed a systematic approach for achieving fairness. The core idea is to reduce fair classification to a series of cost-sensitive classification problems. The solutions obtained yield a randomized classifier with the lowest empirical error while adhering to specified constraints. For fair regression, ^[80] introduced a general scheme wherein the prediction is required to be statistically independent of the protected attribute, and the prediction error for any protected group must remain below a predetermined level.