Auditing Algorithms: Research Methods for Detecting Discrimination on Internet Platforms by Sandvig et al

Measuring Personalization of Web Search by Hannak et al

Summary:

In the first paper authors believe that every algorithm deserves scrutiny as it might be manipulated and discriminatory. They introduce the social scientific audit study used to test for discrimination in housing etc. and propose a similar idea of “algorithm audits” to find out algorithmic bias. They then outline five algorithmic audit designs namely code audit, non-invasive user audit, scraping audit, sock puppet audit and collaborative or crowdsourced audit and discuss the advantages and disadvantages of each. They find the ‘crowdsourced audit technique’ to be most promising.

In the second paper, authors study personalization in algorithms using the “crowdsourced audit” technique described in the first paper. They propose a methodology to measure personalized web results and apply it on 200 mechanical turks and observe that 11.7% of search results show differences due to personalization. Only two factors- being logged in to google account and geographic location leads to measurable personalization. Queries associated with ‘politics’ and ‘companies’ were associated with highest personalization.

Reflection:

Companies have always been altering algorithms to their benefit. Google takes advantage of its market share to promote its services like ‘google maps’. One will almost never find search results containing URLs to MapQuest, Here WeGo and yahoo news. Market monopoly can be a dangerous thing. It kills competition. Who knows if google starts charging for its free services in the future when we all get used to its products.

A case of gender discrimination was found in Linked where in response to search for a female contact name you’ll be prompted male versions of that name. Example: “A search for “Stephanie Williams brings up a prompt asking if the searcher meant to type “Stephen Williams” instead.”[1]. While google example shows an intentional bias which although is not harming the users directly but is killing the market competition, the linkedin incident seems to be an unintentional bias that cropped up in their algorithm since it depended on relative frequencies of words appearing in the queries. So probably ‘Stephan’ was searched more than ‘Stephanie’.  Citing their spokesperson “The search algorithm is guided by relative frequencies of words appearing in past queries and member profiles, it is not anything to do [with] gender.” So authors are right when they say that no algorithm should be considered unbiased.

Some companies are building Tools to detect bias in their AI algorithms like Facebook (Fairness Flow)[2], Microsoft [3] and Accenture[4]. But the problem is that just like their algorithms these tools will be a black box for us. And we will never know if these companies found bias in their algorithms.

Privacy vs personalization/convenience:  Shouldn’t users have the control over their data. Of what they want to share with the companies? Google was reading our mails for almost a decade for personalised advertisements before it stopped that in 2017 [5]. It still reads them though. It knows about our flight schedules, restaurant reservations. My phone number get distributed to so many retailers, I wonder who is selling them this data

In the second paper the authors mention that once the user logs in to one of the google services they are automatically logged-in to all. So does that mean my YouTube search affects my Google search?

According to an article [6] google autocomplete feature is leading to spread of misinformation. The first suggestion that comes up when you type “climate change is” comes out to be “climate change is a hoax”. How is misinformation and conspiracy theories ranking up on these platforms?

Determining bias seems like a very complex problem with online algorithms changing everyday. And there could be multiple dimensions to bias: gender, age, economic status, language, geographical location etc. The collaborative auditing seems to be a good way of collecting data provided it is done systematically and testers are chosen properly. But then again, how many turkers one should hire? Can a few 100 represent the billion population that is using the internet?

[1] https://www.seattletimes.com/business/microsoft/how-linkedins-search-engine-may-reflect-a-bias/

[2] https://qz.com/1268520/facebook-says-it-has-a-tool-to-detect-bias-in-its-artificial-intelligence/

[3] https://www.technologyreview.com/s/611138/microsoft-is-creating-an-oracle-for-catching-biased-ai-algorithms/

[4] https://techcrunch.com/2018/06/09/accenture-wants-to-beat-unfair-ai-with-a-professional-toolkit/

[5] https://variety.com/2017/digital/news/google-gmail-ads-emails-1202477321/

[6] https://www.theguardian.com/technology/2016/dec/16/google-autocomplete-rightwing-bias-algorithm-political-propaganda

Read More

1. Sandvig, Christian, et al. “Auditing algorithms: Research methods for detecting discrimination on internet platforms.” Data and discrimination: converting critical concerns into productive inquiry (2014): 1-23.
2. Hannak, Aniko, et al. “Measuring personalization of web search.” Proceedings of the 22nd international conference on World Wide Web. ACM, 2013.

Both papers deal with the topic of examining personalization and recommendation algorithms that lie at the heart of many online businesses, be it travel booking, real estate, or your plain old web search engine. The first paper “Auditing algorithms” brings up the potential for bias creeping in or even the intentional manipulation of the algorithms to advantage or disadvantage someone. It talks about the need for transparency and proposes that algorithms should be examined via audit study and proposes several methods for doing so. The second paper attempts to identify the triggers of and measure the amount of personalization in a search engine, while controlling for other aspects that could change results but is not relevant to personalization.

I think the first problem one might run into when thinking about attempting an audit study of the algorithms of an enormous entity like Google, Facebook or Youtube is the sheer scale of the task ahead of you. When you are talking about algorithms serving millions, even billions of people across the world, you have algorithms that are working with thousands or tens of thousands of variables and it is working towards finding the optimum values for each individual user. I speculate that slight change in the user’s behavior might set of a chain reaction of variable changes in the algorithm. At this scale, human engineers are no longer in the picture and also the algorithm is evolving on its own (thanks machine learning!) and it is possible that even the people who created the algorithm no longer understand how it works. Why do you think Facebook and Youtube are constantly fighting PR fires? They don’t have as much knowledge or control of their algorithms as they might claim. Even the most direct method of a code audit might see the auditors make some progress before they lose it all because the algorithm changed out from under them. How do you audit an ever shifting algorithm of that much size and complexity? The only thing I can think of is use another algorithm that audits the first algorithm since humans can’t do it at scale. But now you run into the problem of possible bias in the auditor algorithm. It’s turtles all the way down.

Even if we are talking about auditing something of a smaller scale, an audit study is still not a perfect solution because of the possibility of things slipping through the cracks. Linus’s law “Given enough eyeballs, all bugs are shallow” doesn’t really work even when everything is out in the open for scrutiny. OpenSSL was open source and a critical piece of infrastructure but the Heartbleed bug lay there unnoticed for two years regardless of many people looking for bugs. What can we do to improve the audit study methods to catch all instances of bias without allowing the study to become impractically expensive?

Coming to the second paper, I find it fascinating the vast difference in how much the later rank results change compared to rank 1. What I want to know is why are the rank 1 results so relatively stable? Is it simply a matter of having a very high pagerank and being of maximum relevance? Are there cases where a result is hard coded in for search queries (like how you often see a link to wikipedia as the first or second result in many search results)? I think focusing specifically on the rank 1 results would be an interesting experiment. Tracking the search results over a longer period of time and looking at the average time periods between rank 1 results changing and also looking at what kind of search queries see the most volatility in rank 1 results.

Read More

Readings:

[1] Auditing Algorithms: Research Methods for Detecting Discrimination on Internet Platforms

Summary:

The central question that this paper addressed was the mechanisms that are available to scholars and researchers to determine the operation of the algorithms. The authors started talking about the traditional reasons why an audit was carried out, explained how audits were carried out traditionally and why it was ok to cross some ethical borders to find answers for the greater good of the public. They went on to detail how overly restrictive laws (CFAA) and scholarly guidelines are a serious impediment today for a similar study that would not be bound by such laws and guidelines in the 1970’s, consequentially hindering social science researchers from finding answers to the problems they need to solve. Throughout the paper the authors profiled and detailed five varying algorithm audit designs including code audit, noninvasive user audit, scraping audit, sock puppet audit, and collaborative audit or crowdsourced audit.

Reflection/Questions:

Through the entirety of the paper the authors addressed algorithms as something that has conscience and this method of addressing algorithms bothered me, for instance the last question that the author poses  “how do we as a society want these algorithms to behave?”. Usage of the word behave was not apropos according to me and a better fitting word would have been function, so something along the lines “how do we as a society want these algorithms to function?” The authors of this paper also addressed various issues regarding algorithmic transparency that I brought up in my previous blog and in class — ” On many platforms the algorithm designers constantly operate a game of cat-and-mouse with those who would abuse or “game” their algorithm. These adversaries may themselves be criminals (such as spammers or hackers) and aiding them could conceivably be a greater harm than detecting unfair discrimination in the platform itself”. Within the text of the paper the authors contradicted themselves by first saying that audits are carried out to find out trends and not punish any one entity, howbeit later in the paper they say that algorithmic audits on a wide array of algorithms will not be possible and ergo the researchers would have to resort to targeting individual platforms. I disagree that algorithms can incur any sort of bias since biases are based out of emotions, and pre-conceived notions which are a part of human conscience and algorithms don’t have emotions. On that end, let’s say that research finds a specific algorithm on a platform to be biased, who is accountable ?  the company ? the developer/ the developers who created the libraries? the manager of the team?  Lastly, according to me googles “screen science” was perfectly acceptable — one portion of the corporation supporting another portion, just like the concept of a donor baby.

[2 ]  Measuring Personalization of Web Search

Summary:

In this paper the authors detail their methodology for measuring personalization in web searches, apply their methodology to numerous users, and finally dive into cause of personalization on web. The methodology created by the researchers exposed that 11.7% searches were personalized, mainly caused due to geographic location of the user, and the users account being logged in. The method for finding out personalization also controlled for various noise sources, hence delivering more accurate results. The authors acknowledged the drawback in their methodology — which will only identify positive instances of personalization and will not identify absence of personalization.

Reflection/Questions:

Filter bubble’s and media go hand in hand. People consume what they want to consume. Like I have previously said, personalizing search outputs isn’t the evil of all societal problems. According to me it almost seems as if personalization is being associated with manipulation, which is not the same. If search engines do not personalize, the users get frustrated and find a place that will deliver them the content that they want. I would say there are two different types of searches: factual searches, and personal searches. Factual searches include searches which have a factual answer and there is no way that can be manipulated/personalized, however personal searches include things about feelings, products, ideas, perceptions, etc. and these results are personalized, which I think should rightly be.  Authors also write that there is a “possibility that certain information may be unintentionally hidden from users,” which is not a draw back of personalization but reflective and indicative of real life, where  a person is never exposed to all the information on one topic. Howbeit, the big questions I think about personalization are what is the threshold of personalization ?  At what point is the search engine a reflection of our personality and not an algorithm anymore ? At what point does the predictive analysis of searches becomes creepy ?

Read More

Paper 1: Sandvig, Christian, et al. “Auditing algorithms: Research methods for detecting discrimination on internet platforms.” Data and discrimination: converting critical concerns into productive inquiry (2014): 1-23.

Paper 2: Hannak, Aniko, et al. “Measuring personalization of web search.” Proceedings of the 22nd international conference on World Wide Web. ACM, 2013.

It is safe to say that the issue of algorithmic bias and personalization was once celebrated as a “smartness” and high service quality of internet information providers. In a study by Lee et. al. [1] the research participants attributed algorithms’ fairness and trustworthiness to its perceived efficiency and objectivity. This means reputed and widely acknowledged algorithms such as Google search appear to be more trustworthy. This makes it particularly severe when algorithms give out discriminatory information or make decisions on user’s behalf. Christian et. al.’s paper reviews the implications of discrimination on internet platforms along with the auditory methods that researchers can use to detect its prevalence and effect. Amongst the several methods proposed, Hannak et. al. in paper two utilize sockpuppet audit and crowdsourcing audit methods to measure personalization on Google search. From the review of two papers, it can be said that the bias in algorithms can occur either from the code or from the data.

1. Bias in code can be attributed to financial motives. The examples given in the paper 1 about American Airlines, Google Health and product ranking highlight this fact. But, how is it different from getting store brand at a low price in any of the supermarkets? On surface, both enterprises are utilizing the platform they created to best sell their products. However, the findings in the paper 2 prove that website ranking (either having AA flight ads at the top or having Google service links at the top) is what separates fair algorithms from the unfair algorithm. (Unlike biased information providers Kroger displays its store brand at the same front as other brands). There is a clear difference between change in search rank between AMT results and the control results.
2. Bias in data, I believe, is mainly caused due to the user’s personal history and the dominance of a particular type of information available. Getting similar type of information based on history can lead to echo chambers of ideologies as seen in the previous paper. There is also another type of bias in data that informs algorithms in the form of word embeddings in automatic text processing.  For example, in the paper “Man is to Computer Programmer as Woman is to Homemaker”, [2] Bolukbasi et. al. state that the historical evidence of embeddings of computer science being closer to male names than female names will make search engines rank male computer scientist web pages higher than female scientists. This type of discrimination can not be blamed on a single entity but just the prevalence of biased corpus and the entire human history!

Further, I would like to comment briefly on the other two research design methods suggested in paper 1. Scraping audit can have unwanted consequences. What happens to the data that is scraped and later blocked (or moderated) by the service provider? Recently, Twitter suspended  Alex Jone’s profile but his devoted followers were able to rebuild a fake profile with real tweets based on the data collected from the web crawlers and scrapers. Also, noninvasive user audits, even though completely legal can be ineffective with poor choice of experts.

Finally, given the recent events, it can be valuable to research how algorithms share information across platforms. It is common to see ads of hotels and restaurants on Facebook after booking flight tickets with Google flights. Is “Google personalization” only limited to Google?  

[1] Lee, Min Kyung. “Understanding perception of algorithmic decisions: Fairness, trust, and emotion in response to algorithmic management.” Big Data & Society 5.1 (2018): 2053951718756684.

[2] Bolukbasi, Tolga, et al. “Man is to computer programmer as woman is to homemaker? debiasing word embeddings.” Advances in Neural Information Processing Systems. 2016.

Read More

The amount and sources of information, news, and opinions that we get exposed to every day have significantly increased thanks to online social media platforms. With these platforms serving as a mediator for sharing information between their users, multiple questions arise regarding how their design and function affect the information that reaches an end-user.

In designing an online platform with an overwhelming amount of information flowing every day, it may make sense to design some personalization techniques to optimize the user’s experience. It might also make sense for example for an advertising platform to optimize the display of adds in a way that affects what an end-user see. There are many design goals that may result in the end-user receiving filtered information, however, the lack of transparency of these techniques to end-users, as well as the effect of filtering information on the quality and diversity of the content that reaches a user are significant concerns that need to be addressed.

In their work Exposure to ideologically diverse news and opinion on Facebook, Eytan Bakshy et al. study the factors affecting the diversity of the content that Facebook users are exposed to. They conducted a data-driven approach to analyze the proportion of content from a different ideology versus content from an aligning ideology that a user sees in his Facebook newsfeed. They inferred that the most contributing factors to limiting the diversity of the newsfeed content of a user are the structure of the user’s friends network and what a user chooses to interact with. The study found that the newsfeed ranking algorithm affects the diversity of the content that reaches a user, however, this algorithm is adaptable to the behavior and interactions of the user. From these perspectives, they concluded that “the power to expose oneself to perspectives from the other side in social media lies first and foremost with individuals” as mentioned in the paper.

I agree to some extent with the findings and conclusions of the study discussed above. However, one major concern is the question of to what extent are Facebook users aware of these newsfeed ranking algorithms? Eslami et al. try to answer this critical question in their work “I always assumed that I wasn’t really that close to [her]”: Reasoning about invisible algorithms in the news feed. They conducted a qualitative study that included information from 40 Facebook users about their awareness of the newsfeed curation algorithms. The study showed that the majority are not aware of these algorithms and that there is a large misinterpretation of the effect of these algorithms among users. Although after becoming aware that Facebook controls what they see, a majority of users appreciated the importance of these algorithms, the initial response to knowing that these algorithms exist was highly negative, it also revealed how people are making wrong assumptions when for example not seeing a post from a friend for a while.

I’ll imagine myself as a mediator between users and designers of a Facebook-like social platform, trying to close the gap. I totally agree that every user has the complete right to know how their newsfeed work. And every user should feel that he/she is in full control over what they see and that any hidden algorithm is only helping them to personalize their newsfeed. On the other hand, it is a hard design problem for the platform designers to reveal all their techniques to end-users, simply because the more complex it becomes to use the platform, the more likely users will abandon it to more simple platforms. 

If I imagine being hired to alter the design of a social platform to make users more aware of any hidden techniques. I would start with a very simple message conveyed through an animated video that raises the users’ awareness of how their newsfeed work. This could be by simply saying that “we are working to ensure you the best experience by personalizing your newsfeed, we would appreciate your feedback”. For having a user’s feedback, they could see occasional messages that ask them simple questions like “you’ve been interacting with x recently, to see more posts from x, you can go to settings and set this and that”. After a while, users will become more aware of how to control what they see on their newsfeed. Also taking continuous feedback from users on their satisfaction levels with their experience with the platform will help to improve the design over time.

I understand that it is more complex and challenging to address such a problem and that there may be hundreds of other reasons why there are some hidden algorithms that control what an end-user receives. However, ensuring a higher level of transparency is crucial to the healthiness and user satisfaction with online social platforms.

Read More

PAPERS:

2015. Eslami et al., ““ I always assumed that I wasn’t really that close to [ her ]”: Reasoning about invisible algorithms in the news feed,” 2015.

Bakshy, “Exposure to ideologically diverse news and opinion on Facebook,” vol. 348, no. 6239, pp. 1130–1133, 2015.

SUMMARY:

The first paper deals with Facebook’s ‘algorithm’ and how it’s operations are opaque and yet the power they posess in influencing user’s feeds and thereby virtual interactions and perhaps opinions. Using FeedVis, they do a longitudinal study of the algorithm and how the awareness of the algorithm influences the way people interact with Facebook and feel satisfied or otherwise.

The second platform deals with the algorithm’s nature to make the platform, an echo chamber owing to the homophilly patterns among users with a certain ideologies or opinions.

REFLECTION:

• The paper itself mentions and I agree, that the results from the FeedVis study were logitudinal and more knowledge about the user patterrns could be achieved by taking into account ethnography and how the algorithm works for different for different users.
• Another factor which the paper briefly touches upon is that the users try to speculate about this ‘opaque’ algorithm and find ways to understand and ‘hack’ into the algorithms and thus the respective feeds of their followers.
  • One such example is the entire YouTube and Instagram community trying to constantly figure out the algorithms of the respected platforms and adjusting their activities online in accordance with those.
  • Further, the lack of communication of such algorithms often demotes the feeling of a community among users and thereby affecting the user ‘patriotism’ towards the platform.
    • This was observed in the YouTube Demonitization case where several YouTubers, due to lack of effective communication from YouTube, felt less important and thus changed their online activities.
• Furthermore, I would have liked if these studies were conducted in today’s times, mentioning Dark Posts or Unpublished Posts and how the ‘algorithm’ treats them and how is bolsters the homophily (often political) in users.
• The use of Dark Posts is very unethical as it promotes the ‘echo chambers’ on social media sites. Not only that, the users, differing in ideologies to a certain demographic will not even see the post organically due to the “unpublished ness” of the post. Allegedly, even a link to that post will not take a user to that post if the user’s interests have been profiled different from the targeted audience of the Dark Post. Dark Posts can not only be used for targeted marketing but also for certain other unethical areas. [1]

[1] – http://adage.com/article/digital/facebook-drag-dark-posts-light-election/311066/

Read More

1. “I always assumed that I wasn’t really that close to [her]: Reasoning about Invisible Algorithms in News Feeds.”
2. “Exposure to ideologically diverse news and opinion on Facebook.”

Summary:

The first paper follows a very deep qualitative approach studying the awareness of Facebook users towards the algorithms that curates their news feed. And how people react when they know that their news feed is not random or inclusive. They developed a system – FeedVis – that shows users stories from their families News feed and allows users to control their own news feed. They try to answer three questions:

• How aware are the users about their news feed curation algorithm?
  • 5 % were unaware of such algorithm
  • 5% were aware due to different reasons (inductive and deductive)
• What is their reaction when they know about it? Do they prefer the old curated algorithm or the output generated by FeedVis?
• How did the participation of this study affect their usage of Facebook?

The second paper studies data from more than 10 million Facebook users to study what factors affect the nature and ideology of news we receive in our news feed on Facebook. They defined three features that would affect our news feed: 1- User interaction (e.g. clicks) with the shown news, 2- friends’ network shared news and its diversity, and 3- algorithmically ranked news by Facebook news curation algorithm.

They found that what shapes our news feed is what we select to choose and interact with. That might trap us into echo-champers.

Reflection:

• It is amazing how such algorithms could alter people’s feeling and ideas. Some participants lack self-confidence just because nobody reacted to their posts. Awareness of such algorithms increased their postings and interaction with Facebook knowing that nobody react to their posts was due to the curation algorithm
• The authors might do further analysis about the similarities and differences, the backgrounds and beliefs of each participant and their friends who got their stories appear in the news feed. This analysis might help answer a few questions about the news feed curation algorithms of Facebook:
  • Does Facebook really connects people? Or creates more closed communities of common interests and backgrounds?
  • How much those algorithms contribute into increasing polarization? And the possibility to design new tools to alleviate it?
• The second paper answered many of the questions raised from the first one, it highlights the reasons and factors that influence the algorithm in the first paper, which is our own choices and interaction with what is displayed in our news feed. We are the ones who – indirectly – direct our news ranking algorithm, I believe our news feed is just a reflection of our ideology and interests.

• I think the Facebook news feed curation algorithm should be altered in order to alleviate the polarization of its users, creating more diverse interactive healthier environment instead of being trapped in closed-minded separated communities (or echo chambers as the authors call)

Read More

Readings Assigned

1. Bakshy, Eytan et al. – “Exposure to ideologically diverse news and opinion on Facebook” – Published in http://science.sciencemag.org Science 05 Jun 2015:
   348, Issue 6239, pp. 1130-1132 DOI: 10.1126/science.aaa1160
2. Eslami, Motahhare et al.- “I always assumed that I wasn’t really that close to [her]”: Reasoning about invisible algorithms in the news feed” – Proceedings of CHI’15 (153-162)

Summary

The article and the paper both discuss how the Facebook notification feed is curated by an algorithm. The first article published in the science magazine talks about how a user of Facebook is exposed to diverse amount of news and opinion. The authors deduced that social median in fact do expose the users to ideologically diverse viewpoints. The authors of the second paper conducted a Facebook news feed case study comprising of 40 Facebook users. The users were from a diverse social as well as economic background. They classified the users as “aware” or “unaware” based on their knowledge of Facebook curation of the news feed which is based on an algorithm. The theme of this week’s reading was how users are influenced by the notification feed and how they react when knowing about the existence of the algorithm, their satisfaction level after being given a chance to see their unfiltered feed. Overall, it was found that the users were ultimately satisfied about the way the news feed was curated by the algorithm. All of them became more aware of the algorithm. Th experiment changed the the way they used and interacted with posts as that decision was a much informed one.

Reflection

 It is true that computer algorithm exposes us to content which sometimes do influence our ideology, beliefs and the way we perceive an issue. The only question that comes to my mind is, are algorithms reliable? We know that algorithms are used in almost all things we do on the internet. I read an article on how United Airlines had overbooked a flight [1]. This led to more people possessing tickets than the number of seats available. Upon discovery of this, a passenger was removed forcibly from the plane. The reason that the airline company had given was that an algorithm had sorted through the passenger list and took in some parameters like the price of the ticket, if they were frequent flyer and the time of their check in. It had thus given the output that the passenger who was removed was one who was “least valuable” to them.

Additionally, algorithms are used profusely in each and every aspects of the internet. Social media’s news feeds are curated. The one thing that companies could potentially do to improve user awareness of the algorithmic existence is to inform users about them. I do not mean the endless “terms and conditions”. What I do mean is, like Facebook reminds one of all the memories, birthday, they can remind or notify about the algorithm. Since social media users are varied in education status and background, and that not all are from “computer science” background, it is the responsibility of the company to make sure users are “aware”.

Moreover, they can also provide more flexibility to users to filter the notification. I know that these are already in place but are a bit ambiguous. It depends on users indicating their preference against each post. However, such filtering should be made more user friendly and easy. Similar to how we filter amazon search results, something like that can be implemented in the homepage globally, not just against each post. It can be chronologically by default and customization on demand. Facebook in particular starts showing posts related to what we have recently liked or visited. This generally leads to the feed being monopolized with certain posts and that generally is one of the main reasons I am repelled by the platform. Advanced filtering setting could have these parameters as well to help users even more and allow users to customize rather than the algorithm choosing for us.

[1] https://99percentinvisible.org/episode/the-age-of-the-algorithm/

Read More

Papers –

[1] “I always assumed that I wasn’t really that close to [her]”: Reasoning about invisible algorithms in the news feed – Eslami et. al.

[2] Exposure to ideologically diverse news and opinion on Facebook – Bakshy et. al.

Summary –

Read More

PAPERS:

• Bakshy, “Exposure to ideologically diverse news and opinion on Facebook,” vol. 348, no. 6239, pp. 1130–1133, 2015.
• M. Eslami et al., ““ I always assumed that I wasn’t really that close to [ her ]”: Reasoning about invisible algorithms in the news feed,” 2015.

SUMMARY:

Paper 1:

The question that was the central part of the paper was “How do [these] online networks influence exposure to perspectives that cut across ideological lines?” for which de-identified data of 10.1 million U.S. Facebook users were measured for their ideological homophily in friend networks. The influence of ideologically discordant and the relationship with the heterogeneity of friends with such data led the authors to conclude that “individuals’ choices played a stronger role in limiting exposure to cross-cutting content.”

The comparisons and observations were captured in:

• comparing the logical diversity of the broad set of news and opinion shared on Facebook with that shared by individuals’ friend networks
• comparing this with the subset of stories that appear in individuals’ algorithmically ranked News Feeds
• observing what information individuals choose to consume, given exposure on News Feed.

A point of interest as a result of the study was the suggestion that the power to expose oneself to perspectives from the other side (liberal or conservative) in social media lies first and foremost with individuals.

Paper 2:

The objective of the paper was to find “whether it is useful to give users insight into these [social media] algorithms’ existence or functionality and how such insight might affect their experience”. The development of a Facebook application called FeedVis for this purpose helped them answer three questions:

• How aware are users of the News Feed curation algorithm and what factors are associated with this awareness?
• How do users evaluate the curation of their News Feed when shown the algorithm outputs? Given the opportunity to alter the outputs, how do users’ preferred outputs compare to the algorithm’s?
• How does the knowledge users gain through an algorithm visualization tool transfer to their behavior?

During the study tools of Usability study like think aloud, walkthroughs, questionnaires were employed to extract information from users. The statistical tools of Welch’s test, Chi-square test, Fisher’s exact test helped corroborate findings. The features, both passive and active, that were extracted as a potential explanation for the questions: While all the participants were exposed to the algorithm outputs, why were the majority not aware of the algorithm’s existence? Were there any differences in Facebook usage associated with being aware or unaware of the News Feed manipulation?

REFLECTIONS:

My reflection on this paper might be biased as I am under the impression that the authors of the paper are also stakeholders in the findings resulting in a conflict of interest. I would like to support my impression with a few of the reporting done by the paper:

• The indication or suggestion of individuals choice resulting in the content that one consumes seems to suggest that the algorithm is not controlling the things individuals see but humans indirectly are which is essentially arguing against the second paper we read.
• The limitations as stated by the author make it seem as if the author is leading us to believe in a models findings which are not robust and has the potential to be skewed.

I will acknowledge the fact that the author has a basis for the claims on cross-cutting of data and given a more robust model compensating for all the drawbacks mentioned has the same findings I will be inclined to side with the author’s findings.

The notion of echo chambers and filter bubbles point us to the argument made by the second paper where through a study it shows the need for explainability and the option to choose. This was a paper that I gave a lot of attention to as I feel close to home. I feel that the paper is a proponent for explainable AI. It tries to address the issue of the black box approach most ML and AI algorithms have with even industry leaders only aware of the inputs and outcomes not able to completely reason with the physics or mechanics behind the processing agent or algorithm. As someone who sees the need for Explainability as a requirement to build Interactive AI, I thought the findings of the paper “but obvious” at points. The fact that people expressed anger and concern falls in line with a string of previous findings resulting in the work in [1], [2], [11]–[13], [3]–[10]. Reading through these papers helps one understand the need of the hour.

The paper also approaches the problem from a Human Factors perspective rather than an HCI one which I feel is warranted. I would argue that a textbook approach is not one that is required. I would tangentially propose a new approach for a new field. Expecting one to stick to design principles, analysis techniques that were coined or thought off in an era where the current algorithms were science fiction is ludicrous according to me. We need to approach the analysis of such Human-Centered systems partly with Human Factors, partly psychology and mostly HCI.

I will be really interested in working with developing more understandable AI systems for the layman.

REFERENCES:

[1]        I. John D. Lee and Katrina A. See, University of Iowa, Iowa City, “Trust in Automation: Designing for Appropriate Reliance,” Hum. Factors, vol. 46, no. 1, pp. 50–80, 2004.

[2]        M. T. Ribeiro, S. Singh, and C. Guestrin, “‘Why Should I Trust You?’: Explaining the Predictions of Any Classifier,” pp. 1135–1144, 2016.

[3]        A. Freedy, E. DeVisser, G. Weltman, and N. Coeyman, “Measurement of trust in human-robot collaboration,” in 2007 International Symposium on Collaborative Technologies and Systems, 2007, pp. 106–114.

[4]        M. Hengstler, E. Enkel, and S. Duelli, “Applied artificial intelligence and trust-The case of autonomous vehicles and medical assistance devices,” Technol. Forecast. Soc. Change, vol. 105, pp. 105–120, 2016.

[5]        K. A. Hoff and M. Bashir, “Trust in automation: Integrating empirical evidence on factors that influence trust,” Hum. Factors, vol. 57, no. 3, pp. 407–434, 2015.

[6]        E. J. de Visser et al., “Almost human: Anthropomorphism increases trust resilience in cognitive agents,” J. Exp. Psychol. Appl., vol. 22, no. 3, pp. 331–349, 2016.

[7]        M. T. Dzindolet, S. A. Peterson, R. A. Pomranky, L. G. Pierce, and H. P. Beck, “The role of trust in automation reliance,” Int. J. Hum. Comput. Stud., vol. 58, no. 6, pp. 697–718, 2003.

[8]        L. J. Molnar, L. H. Ryan, A. K. Pradhan, D. W. Eby, R. M. St. Louis, and J. S. Zakrajsek, “Understanding trust and acceptance of automated vehicles: An exploratory simulator study of transfer of control between automated and manual driving,” Transp. Res. Part F Traffic Psychol. Behav., vol. 58, pp. 319–328, Oct. 2018.

[9]        A. Freedy, E. DeVisser, G. Weltman, and N. Coeyman, “Measurement of trust in human-robot collaboration,” in 2007 International Symposium on Collaborative Technologies and Systems, 2007, pp. 106–114.

[10]      T. T. Kessler, C. Larios, T. Walker, V. Yerdon, and P. A. Hancock, “A Comparison of Trust Measures in Human–Robot Interaction Scenarios.”

[11]      M. Lewis, K. Sycara, and P. Walker, “The Role of Trust in Human-Robot Interaction.”

[12]      D. B. Quinn, “Exploring the Efficacy of Social Trust Repair in Human-Automation Interactions.”

[13]      M. Lewis et al., “The Effect of Culture on Trust in Automation: Reliability and Workload,” ACM Trans. Interact. Intell. Syst. ACM Trans. Interact. Intell. Syst. ACM Trans. xxxxxxxx Mon. YYYY, vol. 30, no. x, 2016.

Read More