Summary

The authors believe that in order to ensure fairness in machine learning systems, it is mandatory to have a human in the loop process. In order to identify fairness problems and make improvements, they suppose relying on developers, users, and the general public is an effective way to follow that process. The paper conducts an empirical study with four types of programmatically generated explanations to understand how they impact people’s fairness judgments of ML systems. They try to answer three research questions:

• RQ1 How do different styles of explanation impact fairness judgment of a ML system?
• RQ2 How do individual factors in cognitive style and prior position on algorithmic fairness impact the fairness judgment with regard to different explanations?
• RQ3 What are the benefits and drawbacks of different explanations in supporting fairness judgment of ML systems?

The authors focus on a racial discrimination case study in terms of model unfairness and Case-specific disparate impact. They performed an experiment with 160 Mechanical Turk workers. Their hypothesis proposed that given local explanations focus on justifying a particular case, they should more effectively surface fairness discrepancies between cases. 

 The authors show that: 

• Certain explanations are considered inherently less fair, while others can enhance people’s confidence in the fairness of the algorithm
• Different fairness problems-such as model-wide fairness issues versus case-specific fairness discrepancies-may be more effectively exposed through different styles of explanation
• Individual differences, including prior positions and judgment criteria of algorithmic fairness, impact how people react to different styles of explanation.

Reflection

This is a really informative paper. I like that it had a straightforward hypothesis and chose one existing case study that they evaluated. But I would have loved to see this addressed with judges instead of crowdworkers. They mentioned it in their limitations and I hope that they find enough judges willing to work on a follow-up paper. I believe that they would have insightful knowledge to contribute especially since they practice it. It would give a more meaningful analysis to the case study itself from professionals in the field.

I also wonder how this might scale to different machine learning systems that cover similar racial biases. Having a specific case study makes it harder to generalize even for something in the same domain. But definitely worth investigating since there are so many existing case studies! I also wonder if changing the case study analyzed, we’d notice a difference in the local vs. global explanations patterns in fairness judgement. And how would a mix of both affect the judgement, too. 

Discussion

• What are other ways you would approach this case study?
• What are some explanations that weren’t covered in this study?
• How would you facilitate this study to be performed with judges?
• What are other case studies that you could generalize this to with small changes to the hypothesis?

Read More

Authors: Jonathan Dodge, Q. Vera Liao, Yunfeng Zhang, Rachel K. E. Bellamy, Casey Dugan

Summary

This paper discusses how different types of programmatically generated explanations can impact people’s fairness judgments of ML systems. The authors conduct studies with Mechanical Turk workers by showing them profiles from a recidivism dataset and the explanations for a classifier’s decision. Findings from the paper show that certain explanations can enhance people’s confidence in the fairness of the algorithm, and individual differences, including prior positions and judgment criteria of algorithmic fairness, impact how people react to different styles of explanation.

Reflection

For the sake of the study, the participants were shown only one type of explanation. While that worked for the purpose of this study, there is value in seeing the global and local explanations together. For e.g. the input-influence explanations can highlight the features that is more/less likely to re-offend, and allowing the user to dig deeper into the features by showing a local explanation can help in forming more clarity. There is some scope of building interactive platforms with the “overview first, details on demand” philosophy. It is, therefore, interesting to see the paper discuss about the potentials of a human-in-the-loop workflow.

I agree with the paper that a focus on data oriented explanation has the unintended consequence of shifting blame from the algorithms, which can slow down the “healing process” from the biases we interact with when we use these systems. Re-assessing the “how” explanations i.e. how the decisions were made is the right approach. The Effect of Population and “Structural” Biases on Social Media-based Algorithms – A Case Study in Geolocation Inference Across the Urban-Rural Spectrum by Johnson et al. illustrates how bias can be attributed to the design of algorithms themselves rather than population biases in the underlying data sources.

The paper makes an interesting contribution regarding the participants’ prior beliefs and positions and how that impacts the way they perceive these judgments. In my opinion, as a system developer, it seems like a good option to take a position (obviously, being informed and depends on the task) and advocate for normative explanations, rather than appeasing everyone and reinforcing meaningless biases which could have been avoided otherwise.

Questions

1. Based on Figure 1, what other explanations would you suggest? If you were to pick 2 explanations, which 2 would you pick and why?
2. If you were to design a human-in-the-loop workflow, what sort of input would you seek from the user? Can you outline some high-level feedback data points for a dummy case?
3. Would normative explanations frustrate you if your beliefs didn’t align with the explanations (even though the explanations make perfect sense)? Would you adapt to the explanations? (PS Read about the backfire offer here: https://youarenotsosmart.com/2011/06/10/the-backfire-effect/)

Read More

Summary 

The paper highlights one of the major problems that the current digital world faces, algorithmic bias, and fairness in AI. They point out that often ML models are trained on data that in itself is bias and, therefore, may result in amplification of the existing bias. This often results in people not trusting AI models. This is a good step towards explainable AI and making models more transparent to the user. The authors used a dataset that is used for predicting the risk of re-offending, and the dataset is known to have a racial bias. Global and local explanations were taken into account across four types of explanations styles, namely, demographic-based, sensitivity based, influence based, and case-based. For measuring fairness, they considered racial discrimination and tried to measure case-specific impact. Cognition and an individuals’ prior perception of fairness of algorithms were considered as measures of individual difference factors. Both qualitative and quantitative methods were taken into account during the evaluation. They concluded that a general solution is not possible but depends on the user profile and fairness issues. 

Reflection 

The paper by Dodge et al. is a commendable effort towards making algorithms and their processing more clear to humans. They take into account not only the algorithmic fairness but also the humans’ perception of the algorithm, various fairness problems, and individual differences in their experiments. The paper was an interesting read, but a better display of results would make it easier for the readers to comprehend. 

In the model fairness section, they are considering fairness in terms of racial discrimination. Later in the paper, they do mention that the re-offending prediction classifier has features such as age included. Additionally, features like gender might play an important role too. It would be interesting to see how age and other features as a fairness issue perform on maybe other datasets where such biases are dominant. 

The authors mentioned that a general solution is not possible to be developed. However, is it possible for the solution to be domain-specific? For example, if we change the dataset to include other features for fairness, we should be able to plug in the new data without having to change the model.

The study was done using crowd workers and not domain experts who are well knowledgeable with the jargon and are used to being unbiased. Humans are prone to be biased with/without intentions. However, people who are in the legal paradigm like judges, attorneys, paralegals, court reporters, law enforcement officers are more likely to be impartial because either they are under oath or years’ of practice and training in an unbiased setup. So, including them in the evaluation and utilizing them as expert crowd workers might yield better results.

Questions

1. A general solution for a domain rather than one size fits all?
2. Only racial discrimination is considered as a fairness issue. Are other factors only used as a feature to the classifier? How would the model perform on a varied dataset with other features like gender as a fairness issue?
3. The authors have used the dataset for the judicial system, and they mentioned their goal was not to study the users. I am curious to know how they anonymize the data, and how was the privacy and security of individuals handled here?

Read More

Summary

Dodge et al.’s paper “Explaining Models: An Empirical Study of How Explanations Impact Fairness Judgment” presents an empirical study on how people make fairness judgments of machine learning systems and how different styles of explanation impact their judgments. Fairness issues of ML systems attract research interests during recent years. Mitigating the unfairness in ML systems is challenging, which requires the good cooperation of  developers, users, and the general public. The researchers state that how explanations are constructed have an impact on users’ confidence in the systems. To further examine the potential impacts on people’s fairness judgments of ML systems, they conduct empirical experiments involving crowdsourcing workers on four types of programmatically generated explanations (influence, demographic-based, sensitivity, and case-based). Their key findings include: 1) some explanations are considered more fair, while others have negative impact on users’ trust of the algorithm in regards of fairness; 2) varied fairness issues (model-wide fairness and case-specific fairness) can be detected more effectively through an examination of different explanation styles; 3) individual differences (prior positions and judgment criteria of algorithmic fairness) lead to how users react to different styles of explanation. 

Reflection

This paper shines light on a very important fact that bias in ML systems can be detected and mitigated. There is a growing attention to the fairness issues in AI-powered technologies in the machine learning research community. Since ML algorithms are widely used to speed up the decision making process in a variety of domains, beyond achieving good performance, they are expected to produce neutral results. There is no denying the fact that algorithms rely on data, “garbage in, garbage out.” Hence, it is incumbent to feed the unbiased data to these systems upon developers in the first place. In many real-world cases, race is actually not used as an input, however, it correlates to other factors that make predictions biased. That case is not as easy as the cases presented in the paper to detect but still requires effort to be corrected. A question here would be in order to counteract this implicit bias, should race be considered and used to calibrate the relative importance of other factors? 

Besides the bias introduced by data input, there are other factors that need to be taken into consideration to deal with the fairness issues in ML systems. Firstly, machine bias can never be neglected. The term bias in the context of the high-stakes tasks (e.g. future criminal prediction) is very important because a false positive decision could have a destructive impact on a person’s life. This is why when an AI system deals with the human subject (in this case human life), the system must be highly precise and accurate and ideally provide reasonable explanation. Making a person’s life harder to live in a society or impacting badly a person’s life due to a flawed computer model is never acceptable. Secondly, the proprietary model is another concern. One thing should be kept in mind that many high-stacks tasks such as future criminal prediction is a matter of public matter and should be transparent and fair. That does not mean that the ML systems used for those tasks need to be completely public and open. However, I believe there should be a regulatory board of experts who can verify and validate the ML systems. More specifically, the experts can verify and validate the risk factors used in a system so that the factors could be widely accepted. They can also verify and validate the algorithmic techniques used in a system so that the system incorporates less bias. 

Discussion

I think the following questions are worthy of further discussion.

• Besides model unfairness and case-specific disparate impact, are there any other fairness issues?
• What are the benefits and drawbacks of global and local explanations in supporting fairness judgment of AI systems?
• Are there any other style or element of explanations that may impact fairness judgement you can think about?
• If an AI system is not any better than untrained users at predicting recidivism in a fair and accurate way, why do we need the system?

Read More

SUMMARY

The authors of this paper attempt to study the effect explanations of ML systems have in case of fairness judgement. This work attempts to include multiple aspects and heterogeneous standards in making the fairness judgements that go beyond the evaluation of features. In order to perform this task, they utilize four programmatically generated explanations and conduct a study involving over 160 MTurk workers. They consider the impact caused by different explanation styles – global (influence and demographic-based) as well as local (sensitivity and case-based) explanations, fairness issues including model unfairness and case-specific disparate impact, and the impact of individual difference factors such as cognitive style and prior position. They authors utilized the publicly available COMPAS (Correctional Offender Management Profiling for Alternative Sanctions) data set for predicting risk of recidivism which is known to have racial bias. The authors developed a program to generate different explanation versions for a given data point and conducted an online survey style study wherein the participants were made to judge the fairness of a prediction based on a 1 to 7 Likert scale and had to justify the rating given by them.

REFLECTION

I agree that ML systems are often seen as ‘black boxes’ and that this truly does make gauging fairness issues difficult. I believe that this study conducted was indeed very useful in throwing light upon the need for more well-defined fairness judgement methodologies involving humans as well. I feel that the different explanation styles taken into account in this paper – influence, demographic-based, sensitivity, and case-based were good and helped cover various aspects that could contribute in understanding the fairness of the prediction. I found it interesting to learn that the local explanations helped to better understand discrepancies between disparately impacted cases and non-impacted cases whereas the global explanations were more effective in exposing case-specific fairness issues.

I also found interesting to learn that different regions of the feature space may have varied levels of fairness and fairness issues. Having not considered the fairness aspect of my datasets and the impact this would have on the models I build, this made me realize that it would indeed be important to have more fine-grained sampling methods and explanation designs in order to judge the fairness of ML systems.

QUESTIONS

1. The participants involved in this study comprised of 78.8% self-identified Caucasian MTurk workers. Considering that the COMPAS dataset being considered in this study is known to have racial bias, would changing the percentage of the African American workers involved in these studies have altered the results? The study focused on workers living in the US, perhaps knowing the general judgement of people living across the world from multiple races may have also been interesting to study?
2. The authors utilize a logistic regression classifier that is known to be relatively more interpretable. How would a study of this kind extend when it comes to other deep learning systems? Could the programs used to generate explanations be used directly? Has any similar study been performed with these kinds of more complex systems?
3. As part of the limitations of this study, the authors mention that ‘the study was performed with crowd workers, rather than judges who would be the actual users of this type of tool’. How much would the results vary if this study was conducted with judges? Has any follow-up study been conducted?

Read More

The paper provides research on Fairness, Explainable Artificial Intelligence (XAI), and people’s judgment change. The authors introduce a preprocessing method to reduce the bias of a dataset for known bias-inducing attributes. They also show four explanation methods of the classification results: Sensitivity, Input-Influence, Case, and Demographic. Using different combination of the above configurations, AI classifications of the COMPAS data was presented to MTurk workers for feedback. As a result, the paper reports that case-based explanations were often seen as less fair than other explanation methods. The authors also found that sensitivity explanations are the most effective at addressing unfairness. Finally, the paper shows that the evaluator’s position on machine learning heavily impacts his or her reaction to a classifier output and explanations.

When I looked at the paper’s sample sensitivity explanation, it gave me a strong impression that the system was racist. I think many others would have had a similar thought, especially if they do not have enough knowledge about machine learning and regression. Because of this, it concerned me that some people may be lured more towards making the opposite decision than the one that the AI made as a repulsive reaction. This is clearly adding another bias in the opposite direction. I believe an explanatory model should only give helpful information about the model instead of giving bias. Thinking of a possible solution, the authors could have rephrased the same information in a different way. For example, instead of bluntly saying that the classifier would have made a different decision, the system could have reported the probability for each label. This provides the same information but adds less obvious bias. Another solution would be preprocessing the data to not have the bias in the first place like the authors suggested.

I liked the idea of comparing the subject’s prior position to using ML with their judgment of the classifier. This relates to a reflection I made last week, where I stated the possibility that people may make decisions by putting more weight when the model makes a wrong decision. As I have expected, the paper reported that prior positions do in fact make a huge difference in a user’s judgment. Either building more trust with the users or building the software to effectively address both kinds of users would be needed to address this issue.

The followings are the questions I had while reading the paper:

1. Would there be a possibility where preprocessing the data would add bias to the data instead of removing it? What if the attribute that was thought to be unneeded for the classification was actually crucial to the judgment?

2. The authors state that one of the limitations of their study is conducting it with MTurk workers and not the actual users of the software. Do you think this was really a limitation? The attributes used for the classifier and explanations in their experiment seemed general enough for non-professionals to make a meaningful judgment.

3. If you were to design a classifier with an explanation model, which explanation method would you pick? (Out of Sensitivity, Input-Influence, Case, and Demographic) What do you like about the chosen method?

Read More

The paper explores how people make fairness judgments of ML systems and the impact that different explanations can have on these fairness judgments. The paper also explores how providing personalized and adaptive explanations can support such fairness judgments of ML systems. It is extremely important to ensure algorithm fairness and there is a need to consciously work towards avoiding the risk of amplifying existing biases. In this context, providing explanations can be beneficial in two aspects, not only do they help in providing implementation details which would otherwise be a “black box” to a user, but they also facilitate better human-in-the-loop experiences by enabling people to identify fairness issues. The COMPAS recidivism data was utilized for the study and four different explanations styles were examined: input-influence based, demographic-based, sensitivity-based, and case-based. Through the study, it is highlighted that there is no one-size-fits-all solution for an effective explanation. The dataset, context, kinds of fairness issues, and user profiles vary and need to be addressed individually. The paper proposes providing hybrid explanations as a solution to address this problem thereby providing both an overview of the ML model and information about specific cases to help aid accurate fairness judgment.

While there has been a lot of research focus on developing non-discriminatory ML algorithms, this paper specifically deals with the human aspect which is necessary to identify and remedy fairness issues. I feel that this is equally important and is often overlooked. It was interesting to note that they auto-generated the explanations, unlike previous studies. 

With respect to the different explanation styles used, I found the sensitivity-based explanation particularly interesting since it clearly shows the difference in the prediction result if certain attributes were modified. According to me, this form of explanation, out of the four proposed, is extremely effective in bringing out any bias that may be present in the ML system.

I felt that the input-influence based explanation was also effective since it had the +/- markers corresponding to features that match the particular case and this gives the users a clearer picture of which attributes specifically influenced the result thereby providing the implementation details to a certain extent.

The study results documents various insights from participants, and I found some of them to be extremely fascinating. While some believed that certain predictions were biased, others found it normal for that verdict to be predicted. It truly captured the diversity in opinions and perspectives of the same ML system based on the different explanations provided.

1. Through this study, it is revealed that the perception of bias is not uniform and is extremely subjective. Given this lack of agreement on the definition of moral concepts, how can a truly unbiased ML system be achieved?
2. What are some practices that can be followed by ML model developers to ensure that the bias in the input dataset is identified and removed?
3. Apart from gender-bias and ethnic-bias, what are some other prevalent biases in existing ML systems that need to be eradicated?

Read More

This paper mainly explores the injustice of the results of machine learning. These injustices are usually reflected in gender and race, so in order to make the results of machine learning better serve people, the author of the paper conducted an empirical study with four types of programmatically generated explanations to understand how they impact people’s fairness judgments of ML systems. In the experiment, these four interpretations have different characteristics, and after the experiment, the author has the following findings:

1. Some interpretations are inherently considered unfair, while others can increase people’s confidence in the fairness of the algorithm;
2. Different interpretations can more effectively expose different fairness issues, such as the model-wide fairness issue and the fairness difference of specific cases.
3. There are differences between people, different people have different positions, and the perspective of understanding things will affect people’s response to different interpretation styles.

In the end, the authors obtained that in order to make the results of machine learning generally fair, in different situations, different corrections are needed and differences between people must be taken into account.

Reflection：

In another class this semester, the teacher gave three reading materials on the results of machine learning and increased discrimination. In the discussion of those three articles, I remember that most students thought that the reason for discrimination should not be Is the inaccuracy of the algorithm or model, and I even think that machine learning is to objectively analyze things and display the results, and the main reason that people feel uncomfortable and even feel immoral in the face of the results is that people are not willing to face these results. It is often difficult for people to have a clear understanding of the whole picture of things, and when these unnoticed places are moved to the table, people will be shocked or even condemn others, but it is difficult to really think about the cause of things. But after reading this paper, I think my previous understanding was narrow: First, the results of the algorithm and the interpretation of the results must be wrong and discriminatory in some cases. So only if we resolve this discrimination can the results of machine learning be able to better serve people. At the same time, I also agree with the ideas and conclusions in the article. Different interpretation methods and different emphasis will indeed affect the fairness of interpretation. All the prerequisites to eliminate injustices are to understand the causes of these injustices. At the same time, I think the main solution to eliminate injustice is still on the researcher. Reason why I think computer is fascinating is it can always keep things rational and objective to deal with problems. People’s response to different results and the influence of different people on different model predictions are the key to eliminating this injustice. Of course, I think people will think that part of the cause of injustice is also the injustice of our own society. When people think that the results of machine learning carry discrimination based on race, sex, religion, etc., should we think about this discrimination itself, should we pay more attention to gender equality, ethnic equality and how to make the results look better.

Question:

1. Do you think that this unfairness is more because the results of machine learning mislead people or it is existed in people’s society for a long time.
2. The article proposes that in order to get more fair results, more people need to be considered, what changes should users make.
3. How to combine the points of different machine learning explanations to create a fairer explanation.

Read More

A big concern with the usage of current ML systems is the issue of fairness and bias when making their decisions. Bias can creep into ML decisions through either the design of the algorithm or through training datasets that are labeled in a way to bias against certain kinds of things. The example used in this paper is the bias against African Americans in an ML system used by judges to predict the probability of a person re-offending after committing a crime. Fairness is hard to judge when ML systems are black boxes so this paper proposes that if ML systems expose reasons behind the decisions (i.e. the idea of explainable AI), a better judgement of the fairness of the decision can be made by the user. To this end, this paper examines the effect of four of different kinds of explanations of the ML decisions on people’s judgements of the fairness of that decision.

I believe this is a very timely and necessary paper in these times, with ML systems being used more and more for sensitive and life changing decisions. It is probably impossible to stop people from adopting these systems so the next best thing is making explainability of the ML decisions mandatory, so people can see and judge if there was potentially bias in the ML system’s decisions. It is interesting that people were mostly able to perceive that there were fairness issues in the raw data. You would think that that would be hard but the generated explanations may have worked well enough to help with that (though I do wish they could’ve shown an example comparing a raw data point and a processed data point that showed how their pre-processing cleaned things). I did wonder why they didn’t show confidence levels to the users in the evaluation, but their explanation that it was something they could not control for makes sense. People could have different reactions to confidence levels, some thinking that anything less than a 100% is insufficient, while others thinking that 51% is good enough. So keeping it out is a limiting but is logical.

1. What other kinds of generated explanations could be beneficial, outside of the ones used in the paper?
2. Checking for racial bias is an important case for fair AI. In what other areas is fairness and bias correction in AI critical?
3. What would be ways that you could mitigate any inherent racial bias of the users who are using explainable AI, when they are making their decisions?

Read More

Summary

            Dodge et al. cover a terribly important issue with artificial intelligence programs and biases from historical datasets, and how to mitigate the inherent racism or other biases within. They also work understand how to better communicate why AIs reach the recommendations they do and how. In an experiment, they look at communicating outcomes from a known biased ML model for predicting recidivism amongst released prisoners called COMPAS. They cleaned the ML model to make race less impactful to the final decision, and then produced 4 ways of explaining the result of the model to 160 mTurk workers: Sensitivity, Input-influence, Case, and demographic. “Input” emphasizes how much each input affected the results, “Demographic” describes how each demographic affects the results, “Sensitivity” shows what flipped demographics would have changed the results, and “Case” finds the most similar cases and details those results. They found that local-based explanations (case and sensitivity) had the largest impact on perceived fairness.

Personal Reflection

This study was pretty interesting to me based on it actually trying to adjust for the biases of input data as well as understanding how to better convey insights from less-biases systems. I am still unsure that the authors removed all bias from the COMPAS system but seeing that they did lower the coefficient significantly shows that it was working on it. In this vein, the paper made me want to read the paper they cited as how they could mitigate biases in these algorithms.

I found their various methods on how to communicate how the algorithm came to its recommendation to be rather incisive. I wasn’t surprised that people found that when the sensitivity explanation said that if the individual’s race was flipped the decision would be flipped lead to more perceived issues with the ML decision. That method of communication seems to lead people to see issues with the dataset more easily in general.

The last notable part of the experiment is that they didn’t give a confidence value for each case – they stated that they could not control for it and so did not present it to participants. That seems like an important part of making a decision based on the algorithm. If the algorithm is really on the fence, but has to recommend one way or the other, it might make it easier to state that the algorithm is biased.

Questions

1. Would removing the race (or other non-controllable biases) coefficient altogether affect the results too much? Is there merit in zero-ing out the coefficient of these factors?
2. Having an attention check in the mTurk workflow is, in itself, not surprising. However, the fact that all of the crowdworkers passed the check is surprising. What does this mean for other work that ignores a subset of data assuming the crowdworkers weren’t paying attention? (Like the paper last week that ignored the lowest quartile of results)
3.  What combination of the four different types would be most effective? If you presented more than one type, would it have affected the results?
4. Do you think showing the confidence value for the algorithm would impact the results significantly?  Why or why not?

Read More