260 lines
9.4 KiB
Markdown
260 lines
9.4 KiB
Markdown
file:: [Mehrabi et al. - 2022 - A Survey on Bias and Fairness in Machine Learning.pdf](file://C:\Users\david\Zotero\/storage/VHPDAFE3/Mehrabi et al. - 2022 - A Survey on Bias and Fairness in Machine Learning.pdf)
|
|
file-path:: file://C:\Users\david\Zotero\/storage/VHPDAFE3/Mehrabi et al. - 2022 - A Survey on Bias and Fairness in Machine Learning.pdf
|
|
|
|
- sensitive environments
|
|
ls-type:: annotation
|
|
hl-page:: 1
|
|
hl-color:: green
|
|
id:: 64490503-eb8e-4a42-9fdd-4b4b13e6bf2c
|
|
- do not reflect discriminatory behavior
|
|
ls-type:: annotation
|
|
hl-page:: 1
|
|
hl-color:: green
|
|
id:: 64490510-5fc1-4bbe-aecd-c10ea23f5ebb
|
|
- biases in various ways
|
|
ls-type:: annotation
|
|
hl-page:: 1
|
|
hl-color:: purple
|
|
id:: 6449053d-92ae-4175-9296-9916a34be365
|
|
- different sources of biases that can affect AI applications
|
|
ls-type:: annotation
|
|
hl-page:: 1
|
|
hl-color:: green
|
|
id:: 6449054a-bea4-4b8c-9878-d490a823bb9f
|
|
- Algorithms make movie recommendations, suggest products to buy, and who to date
|
|
ls-type:: annotation
|
|
hl-page:: 1
|
|
hl-color:: green
|
|
id:: 6449095f-7a46-4c26-a748-4599069e22be
|
|
- However, like people, algorithms are vulnerable to biases that render their decisions “unfair”
|
|
ls-type:: annotation
|
|
hl-page:: 2
|
|
hl-color:: purple
|
|
id:: 64490976-44e2-4660-8528-67fc80215dee
|
|
- absence of any prejudice or favoritism toward an individual or group based on their inherent or acquired characteristics.
|
|
ls-type:: annotation
|
|
hl-page:: 2
|
|
hl-color:: purple
|
|
id:: 64490984-7bb9-4f37-acc7-a0e9b92755af
|
|
- skewed toward a particular group of people.
|
|
ls-type:: annotation
|
|
hl-page:: 2
|
|
hl-color:: purple
|
|
id:: 64490a02-fdc1-4cab-8167-b80efe1fc542
|
|
- Correctional Offender Management Profiling for Alternative Sanctions (COMPAS)
|
|
ls-type:: annotation
|
|
hl-page:: 2
|
|
hl-color:: blue
|
|
id:: 64490a19-3f57-414f-8501-a09108f3bfae
|
|
- risk of a person to recommit another crime.
|
|
ls-type:: annotation
|
|
hl-page:: 2
|
|
hl-color:: blue
|
|
id:: 64490a28-9c75-4935-9c44-aea5a3894fcc
|
|
- release an offender or to keep him or her in prison
|
|
ls-type:: annotation
|
|
hl-page:: 2
|
|
hl-color:: green
|
|
id:: 64490a4f-d2f4-4c67-a78f-e93b13f3fb5b
|
|
- facial recognition software in digital cameras that overpredicts Asians as blinking
|
|
ls-type:: annotation
|
|
hl-page:: 2
|
|
hl-color:: green
|
|
id:: 64490a70-059e-4848-8dc8-46af5825245f
|
|
- we observe that biased algorithmic outcomes might impact user experience, thus generating a feedback loop between data, algorithms, and users that can perpetuate and even amplify existing sources of bias
|
|
ls-type:: annotation
|
|
hl-page:: 2
|
|
hl-color:: green
|
|
id:: 64490b15-3caf-43ed-8eb9-5a097962942d
|
|
- safety and fairness constraints have become a significant issue for researchers and engineers
|
|
ls-type:: annotation
|
|
hl-page:: 2
|
|
hl-color:: green
|
|
id:: 64490f8c-db21-45dc-b8cd-18daaafcf325
|
|
- All of these applications have a direct effect in our lives and can harm our society if not designed and engineered correctly, that is, with considerations to fairness.
|
|
ls-type:: annotation
|
|
hl-page:: 2
|
|
hl-color:: purple
|
|
id:: 64491040-2bef-4c75-ac2c-24a4375dc08e
|
|
- COMPAS is an exemplar of a discriminatory system
|
|
ls-type:: annotation
|
|
hl-page:: 3
|
|
hl-color:: purple
|
|
id:: 64491098-e69d-4eeb-acee-7218db342e23
|
|
- advertisements promoting jobs in Science, Technology, Engineering, and Math (STEM) fields
|
|
ls-type:: annotation
|
|
hl-page:: 3
|
|
hl-color:: purple
|
|
id:: 644910a6-5b52-4134-a82e-1aa9f1649234
|
|
- gender-neutral way
|
|
ls-type:: annotation
|
|
hl-page:: 3
|
|
hl-color:: green
|
|
id:: 644910af-0e2e-4d33-8d9c-c6ae5726deb2
|
|
- Recommendations as treatments: Debiasing learning and evaluation
|
|
ls-type:: annotation
|
|
hl-page:: 33
|
|
hl-color:: green
|
|
id:: 644912bc-b382-409d-ac70-cba300316797
|
|
- Recommendations as treatments: Debiasing learning and evaluation.
|
|
ls-type:: annotation
|
|
hl-page:: 33
|
|
hl-color:: green
|
|
id:: 644912c1-0949-4551-b63a-435d315935eb
|
|
- Bias in facial recognition systems [124] and recommender systems [136] have also been largely studied and evaluated and in many cases shown to be discriminative towards certain populations and subgroups.
|
|
ls-type:: annotation
|
|
hl-page:: 3
|
|
hl-color:: green
|
|
id:: 644912c9-2445-4c8d-9ec8-44fc6e340fb6
|
|
- where these biases are coming from and what we can do to prevent them.
|
|
ls-type:: annotation
|
|
hl-page:: 3
|
|
hl-color:: green
|
|
id:: 644912d6-377e-443f-aacf-4f7783480eb4
|
|
- widely used commercial risk assessment software
|
|
ls-type:: annotation
|
|
hl-page:: 3
|
|
hl-color:: green
|
|
id:: 644912e2-ee01-4f23-ba71-616ee966c3d0
|
|
- 7 of those were presented to the people in the study
|
|
ls-type:: annotation
|
|
hl-page:: 3
|
|
hl-color:: green
|
|
id:: 644912f4-b09c-408c-9704-e9208d273de7
|
|
- considering fairness constraints is a crucial task while designing and engineering these types of sensitive tools
|
|
ls-type:: annotation
|
|
hl-page:: 3
|
|
hl-color:: green
|
|
id:: 6449131f-8a8a-4b4a-acab-e7ccaa6d3848
|
|
- introducing tools that can assess the amount of fairness in a tool or system
|
|
ls-type:: annotation
|
|
hl-page:: 3
|
|
hl-color:: purple
|
|
id:: 64491347-9137-4aca-a6f7-a4cb336e5b0c
|
|
- Aequitas [132] is a toolkit that lets users to test models with regards to several bias and fairness metrics for different population subgroups.
|
|
ls-type:: annotation
|
|
hl-page:: 3
|
|
hl-color:: green
|
|
id:: 64491361-223a-4d46-b120-9e81f0226d1e
|
|
- AI Fairness 360 (AIF360) is another toolkit developed by IBM to help moving fairness research algorithms into an industrial setting and to create a benchmark for fairness algorithms to get evaluated and an environment for fairness researchers to share their ideas [11]
|
|
ls-type:: annotation
|
|
hl-page:: 3
|
|
hl-color:: green
|
|
id:: 6449137d-1640-497d-86be-184627860674
|
|
- data driven and require data upon which to be trained
|
|
ls-type:: annotation
|
|
hl-page:: 3
|
|
hl-color:: green
|
|
id:: 644913a4-e1c3-4829-aaaa-9d3655d4f169
|
|
- [:span]
|
|
ls-type:: annotation
|
|
hl-page:: 4
|
|
hl-color:: green
|
|
id:: 644913fd-f97d-46cc-94ae-504f5d4cdfbb
|
|
hl-type:: area
|
|
hl-stamp:: 1682510843561
|
|
- Data to Algorithm.
|
|
ls-type:: annotation
|
|
hl-page:: 4
|
|
hl-color:: green
|
|
id:: 6449146f-8bc7-4739-a671-7422bf4416a9
|
|
- Measurement Bias.
|
|
ls-type:: annotation
|
|
hl-page:: 5
|
|
hl-color:: green
|
|
id:: 644914a3-47d7-43ba-86d1-b08a26763f04
|
|
- how we choose, utilize, and measure particular features
|
|
ls-type:: annotation
|
|
hl-page:: 5
|
|
hl-color:: green
|
|
id:: 644914af-b159-48a7-8c8d-52cd8b1604b2
|
|
- Omitted Variable Bias
|
|
ls-type:: annotation
|
|
hl-page:: 5
|
|
hl-color:: green
|
|
id:: 644914ba-6a2d-4f72-83d9-3ce8d8b331e1
|
|
- Representation Bias.
|
|
ls-type:: annotation
|
|
hl-page:: 5
|
|
hl-color:: green
|
|
id:: 644914de-4c28-4e94-a40c-75108b69c8ca
|
|
- Aggregation Bias.
|
|
ls-type:: annotation
|
|
hl-page:: 5
|
|
hl-color:: green
|
|
id:: 64491511-8f92-451b-9c8a-326cb882111f
|
|
- Any general assumptions about subgroups within the population can result in aggregation bias
|
|
ls-type:: annotation
|
|
hl-page:: 5
|
|
hl-color:: green
|
|
id:: 6449153e-a2b6-4e28-89ff-1901dc3789d8
|
|
- The paradox happened as women tended to apply to departments with lower admission rates for both genders.
|
|
ls-type:: annotation
|
|
hl-page:: 5
|
|
hl-color:: purple
|
|
id:: 64491573-f605-4bae-9c42-274f7cc97a56
|
|
- Algorithmic Bias. Algorithmic bias is when the bias is not present in the input data and is added purely by the algorithm
|
|
ls-type:: annotation
|
|
hl-page:: 7
|
|
hl-color:: green
|
|
id:: 644915f2-2826-48dd-992f-3bba88201c20
|
|
- the user interface and through the user itself by imposing his/her self-selected biased behavior and interaction
|
|
ls-type:: annotation
|
|
hl-page:: 7
|
|
hl-color:: green
|
|
id:: 64491601-db94-49e4-bc8d-6ec972778aa3
|
|
- Popularity Bias. Items that are more popular tend to be exposed more.
|
|
ls-type:: annotation
|
|
hl-page:: 7
|
|
hl-color:: green
|
|
id:: 64491628-f150-46c3-a87a-461cd0d66fb0
|
|
- Evaluation bias happens during model evaluation
|
|
ls-type:: annotation
|
|
hl-page:: 8
|
|
hl-color:: green
|
|
id:: 64491646-e0d3-42e6-bd6d-cc3494af166f
|
|
- Historical bias is the already existing bias and socio-technical issues in the world and can seep into from the data generation process even given a perfect sampling and feature selection
|
|
ls-type:: annotation
|
|
hl-page:: 8
|
|
hl-color:: green
|
|
id:: 64491698-3add-4aaa-aa27-b94c8e8da9a6
|
|
- Population bias arises when statistics, demographics, representatives, and user characteristics are different in the user population of the platform from the original target population
|
|
ls-type:: annotation
|
|
hl-page:: 8
|
|
hl-color:: green
|
|
id:: 644916a0-2631-4fbb-8a55-2d49ec2c60dc
|
|
- feedback loop is not only existent between the data and the algorithm, but also between the algorithms and user interaction
|
|
ls-type:: annotation
|
|
hl-page:: 9
|
|
hl-color:: green
|
|
id:: 644916c0-f9ea-451d-8f18-266efb0f996f
|
|
- Examples of Bias in Machine Learning Data
|
|
ls-type:: annotation
|
|
hl-page:: 9
|
|
hl-color:: green
|
|
id:: 64491720-a503-4bf0-aa89-1e711dfd21f7
|
|
- dark-skinned females
|
|
ls-type:: annotation
|
|
hl-page:: 9
|
|
hl-color:: green
|
|
id:: 64491748-23d6-4475-96b9-407524f0e7ec
|
|
- Similar to bias, discrimination is also a source of unfairness
|
|
ls-type:: annotation
|
|
hl-page:: 10
|
|
hl-color:: green
|
|
id:: 64491956-8fb9-45f0-8472-81aaafe147f5
|
|
- Emergent Bias.
|
|
ls-type:: annotation
|
|
hl-page:: 8
|
|
hl-color:: purple
|
|
id:: 64491a84-f0c7-4bbc-af7c-2c8fc605f702
|
|
- METHODS FOR FAIR MACHINE LEARNING
|
|
ls-type:: annotation
|
|
hl-page:: 13
|
|
hl-color:: purple
|
|
id:: 644a347e-bcf2-4675-8d5e-6e86cbc7b9d9
|
|
- ddress bias in artificial intelligence to achieve fairness
|
|
ls-type:: annotation
|
|
hl-page:: 13
|
|
hl-color:: green
|
|
id:: 644a3492-c375-4549-a723-a59b10f32c80 |