Glimpse of different types of bias in machine learning
Artificial intelligence and machine learning tools are becoming more and more popular in today’s world in many fields.
Generally, ML algorithms work from experiences and then apply them to new data. However, “biased data” problems may affect information gathering or model creation processes, leading to undesirable consequences.
What is the meaning of bias in machine learning?
Tom Mitchell coined the word bias in 1980 in his paper titled “The Need for Biases in Learning Generalizations”. The concept behind bias was to give certain features more weight to generalize better over a broader dataset with several other attributes. Bias in machine learning will help us generalize our models and make them less sensitive to a single data point.
In other words, AI bias is a machine learning algorithm performance paradox that is caused by discriminatory assertions or prejudices in training data during the algorithm development process.
What are the different types of bias?
In their paper named “A Framework for Understanding Unintended Consequences of Machine Learning”, Suresh and Guttag represented bias in 6 different types:
1: Historical bias: it emerges if there is a misalignment between the environment and the priorities to be determined and propagated. It is normative about the world’s state and exists even when there is a perfect set of samples and features collection. For example, in 2018, the Image Search results for “”CEO”” displayed only 5% of women.
— Of course, the data is a perfect reflection of the world, but this might harm some people —
2: Representation bias: it occurs when identifying and sampling a population. In other words, when a population is not well represented inside the data, there is a massive chance that it will fail to generalize well. Datasets collected by mobile applications, for example, may not give a good illustration of poor people or older groups because they are less likely to own smartphones. Let’s take another example: Shankar et al. (2017) showed that the results of a classifier trained on ImageNet are bad for some groups. This poor result is due to the lack of images for those groups. We know most of the ImageNet data were taken in the United States, North America, or Western Europe. Only 1 to 2.1% of the images come from China and India.
3: Measurements bias: It appears when data collected for training varies from data collected in the real world or inaccurate measurements resulting in data distortion. Image recognition datasets are a clear example of this bias. Most of the time the training data is collected with one form of camera while the output data is collected with a different type of camera.Inconsistent annotation during the data labeling stage of a project may also induce measurement bias.
4: Aggregation bias: It occurs when different groups are improperly mixed during model creation. The population of interest in many applications is heterogeneous, and a single model is unlikely to fit all subgroups. This form of bias can result in a not best model for any group or tailored to the dominant population if combined with representation bias. According to Herman and Cohen’s research, HbA1c levels commonly used to diagnose and control diabetes vary in complex ways across ethnicities and genders.
5: Evaluation bias: It happens during the evaluation or iteration process. It usually occurs when the testing or external target populations do not characterize the user population’s different parts. Evaluation bias may also be caused by the use of inadequate metrics for the model’s intended use.
6: Deployment bias: It happens when there is a disparity between the problem that a model is built to solve and how it is currently used. In other words, if a model is designed for a specific purpose and that task is not carried out after deployment. There is no assurance that good evaluation performance will carry over.
What is racial bias?
Machines are often assumed to be neutral, but they are not. In AI systems sold by tech giants including IBM, Microsoft, and Amazon, significant gender and racial bias can be noticed. When data is skewed in favor of some demographics, this is known as racial bias. Face recognition and automatic speech recognition technology, for example, struggle to identify people of color.
Overall, machines are not fair in their judgment as many biases can appear. However, machine learning biases can only be caused by human biases; therefore, it is essential to eliminate such prejudices from the data collection.
References
A Framework for Understanding Unintended Consequences of Machine Learning
https://arxiv.org/pdf/1901.10002.pdf
7 Types of Data Bias in Machine Learning
https://lionbridge.ai/articles/7-types-of-data-bias-in-machine-learning/
Tackling bias in artificial intelligence (and in humans)
