Author: Nate TeBlunthuis

Posted on

Effects of Algorithmic Flagging on Fairness: Quasi-experimental Evidence from Wikipedia

Many online platforms are adopting machine learning as a tool to maintain order and high quality information in the face of massive influxes of of user generated content. Of course, machine learning algorithms can be inaccurate, biased or unfair. How do signals from machine learning predictions shape the fairness of online content moderation? How can we measure an algorithmic flagging system’s effects?

In our paper published at CSCW 2021, I (Nate TeBlunthuis) together with Benjamin Mako Hill and Aaron Halfaker analyzed the RCFilters system: an add-on to Wikipedia that highlights and filters edits that a machine learning algorithm called ORES identifies as likely to be damaging to Wikipedia. This system has been deployed on large Wikipedia language editions and is similar to other algorithmic flagging systems that are becoming increasingly widespread. Our work measures the causal effect of being flagged in the RCFilters user interface.

Screenshot of Wikipedia edit metadata on Special:RecentChanges with RCFilters enabled. Highlighted edits with a colored circle to the left side of other metadata are flagged by ORES. Different circle and highlight colors (white, yellow, orange, and red in the figure) correspond to different levels of confidence that the edit is damaging. RCFilters does not specifically flag edits by new accounts or unregistered editors, but does support filtering changes by editor types.

Our work takes advantage of the fact that RCFilters, like many algorithmic flagging systems, create discontinuities in the relationship between the probability that a moderator should take action and whether a moderator actually does. This happens because the output of machine learning systems like ORES is typically a continuous score (in RCFilters, an estimated probability that a Wikipedia edit is damaging), while the flags (in RCFilters, the yellow, orange, or red highlights) are either on or off and are triggered when the score crosses some arbitrary threshold. As a result, edits slightly above the threshold are both more visible to moderators and appear more likely to be damaging than edits slightly below. Even though edits on either side of the threshold have virtually the same likelihood of truly being damaging, the flagged edits are substantially more likely to be reverted. This fact lets us use a method called regression discontinuity to make causal estimates of the effect of being flagged in RCFilters.

Charts showing the probability that an edit will be reverted as function of ORES scores in the neighborhood of the discontinuous threshold that triggers the RCfilters flag. The jump in the increase in reversion chances is larger for registered editors compared to unregistered editors at both thresholds.

To understand how this system may effect the fairness of Wikipedia moderation, we estimate the effects of flagging on edits on different groups of editors. Comparing the magnitude of these estimates lets us measure how flagging is associated with several different definitions of fairness. Surprisingly, we found evidence that these flags improved fairness for categories of editors that have been widely perceived as troublesome—particularly unregistered (anonymous) editors. This occurred because flagging has a much stronger effect on edits by the registered than on edits by the unregistered.

We believe that our results are driven by the fact algorithmic flags are especially helpful for finding damage that can’t be easily detected otherwise. Wikipedia moderators can see the editor’s registration status in the recent changes, watchlists, and edit history. Because unregistered editors are often troublesome, Wikipedia moderators’ attention is often focused on their contributions, with or without algorithmic flags. Algorithmic flags make damage by registered editors (in addition to unregistered editors) much more detectable to moderators and so help moderators focus on damage overall, not just damage by suspicious editors. As a result, the algorithmic flagging system decreases the bias that moderators have against unregistered editors.

This finding is particularly surprising because the ORES algorithm we analyzed was itself demonstrably biased against unregistered editors (i.e., the algorithm tended to greatly overestimate the probability that edits by these editors were damaging). Despite the fact that the algorithms were biased, their introduction could still lead to less biased outcomes overall.

Our work shows that although it is important to design predictive algorithms to not have such biases, it is equally important to study fairness at the level of the broader sociotechnical system. Since we first published a preprint of our paper, a followup piece by Leijie Wang and Haiyi Zhu replicated much of our work and showed that differences between different Wikipedia communities may be another important factor driving the effect of the system. Overall, this work suggests that social signals and social context can interact with algorithmic signals and together these can influence behavior in important and unexpected ways.

The full citation for the paper is: TeBlunthuis, Nathan, Benjamin Mako Hill, and Aaron Halfaker. 2021. “Effects of Algorithmic Flagging on Fairness: Quasi-Experimental Evidence from Wikipedia.” Proceedings of the ACM on Human-Computer Interaction 5 (CSCW): 56:1-56:27. https://doi.org/10.1145/3449130.

We have also released replication materials for the paper, including all the data and code used to conduct the analysis and compile the paper itself.

Posted on

Community Data Science Collective at CSCW 2021

The Community Data Science Collective is attending the 2021 ACM Conference on Computer-supported Cooperative Work and Social Computing (CSCW).

The conference will feature two new papers by collective students and faculty that were published in the journal Proceedings of the ACM on Human-Computer Interaction: CSCW.

Information on the talks as well as links to the papers are available here (CSCW members are listed in italics):

In addition, Benjamin Mako Hill is a panel co-chair.

Mako, Sohyeon, Jeremy, and Nathan will all be at the conference and so will tons of our social computing friends. Please come and say “Hello” to any of us and introduce yourself if you don’t already know us :)

Posted on

Measuring Wikipedia Article Quality in One Continuous Dimension

Accurately estimating the quality of Wikipedia articles is important task for both researchers and Wikipedia community members. In a forthcoming paper in the Proceedings of the OpenSym 2021, I describe a new method for estimating article quality in Wikipedia in one dimension that builds on the widely used ORES quality model and that improves on the techniques researches have used to incorporate measures of quality into their studies of Wikipedia in the past. I will presenting virtually this week at the OpenSym 2021 conference. OpenSym is free and open to the public this year but requires registration.

Numerous efforts have gone into measuring the quality of Wikipedia articles and the Wikimedia Foundation maintains a machine learning model for measuring article quality on English Wikipedia called the ORES quality model. This model is trained on quality assessments conducted by members of WikiProjects that label articles into hierarchy of quality levels (i.e., stub, start-class, C-class, B-class, Good, Featured) and use boosted decision trees to predict the quality of versions of articles. This model is useful because it can predict the quality of versions of articles that haven’t been assessed. My OpenSym paper (teaser video, full presentation) builds on the ORES quality models to improve measuring Wikipedia article quality in one continuous dimension using ordinal regression models. A 1-D real-valued measure of quality is more granular and easier to use in downstream research. My method also increases the accuracy of quality prediction for units of analysis that are most important for research like articles or revisions and also estimates spacing between different levels of article quality.

Quality scores and predictions of the ordinal regression models. Columns in the grid of charts correspond to the ordinal quality model calibrated to the indicated unit of analysis and rows correspond to sampled articles having the indicated level of quality as assessed by Wikipedians. Each chart shows the histogram of scores, thresholds inferred by the ordinal model with 95% credible intervals colored in gray, and colors indicating when the model makes correct or incorrect predictions. The thresholds are not evenly spaced, especially in revision model and article model that have more weight on lower quality classes. These two models infer that the gaps between Stub and Start and between Start and C-class articles are considerably wider than the gap between C-class and B-class articles.

An important limitation of the ORES quality models is that they do not output a single real-valued quality score. Instead, they output a multinomial distribution of the probability of an article’s quality class. This means that instead of saying “article X has quality 0.3” the model tells you “the probability of article X’s quality class is 0.1 for stub, 0.2 for start-class, 0.5 for C-class, 0.15 for B-class, 0.05 for Good and 0 for featured and the most probable quality class (MPQC) is C.” Using this kind of output in a downstream statistical analysis is kind of messy. For example, it might seem reasonable to use the MPQC as an integer-valued measure of quality, but this throws away information. Suppose ORES says that “the probability of article Y’s quality class is 0.00 for stub, 0.15 for start-class, 0.5 for C-class, 0.2 for B-class, 0.1 for Good, and 0.05 for Featured” According to the ORES quality model, Y probably has greater quality than X. Even though both have a MPQC of C-class, there’s a much greater change for article Y to be B-class or better than for article X. Is there a way to use the ORES probabilities to build a more granular measure of quality that accounts for this difference?

Aaron Halfaker, one of the creators of the ORES system, combined the probabilities into a 1-D score for one of his research projects by taking a weighed sum of the probabilities and assuming that Wikipedia article quality levels are “evenly spaced.” This creates a score out of the quality class probabilities by multiplying each probability by a weight so that probabilities for higher quality levels get more weight. He chose the weights [0,1,2,3,4,5] so an article gets 0 quality points for being a probable stub, 1 for the probability of being start-class, 2 for C-class, and so on with 5 points for Featured. This results in a nice continuous measure of quality that simplifies downstream statistical analysis. A number of others have followed his lead.

But how reasonable is the “evenly spaced” assumption that supports using the weights [0,1,2,3,4,5]? Could there be a way to pick weights to combine the ORES probabilities without using this assumption? My paper explains why ordinal regression is the right tool for the job and proposes a procedure for fitting an ordinal regression model to a sample representative of a unit of analysis like articles or revisions or quality classes that have been labeled by a WikiProject and scored by the ORES quality model. The ordinal regression predicts the quality labels as a function of the ORES scores and in this way finds a good way to combine the ORES scores into a single value. It also infers threshold parameters that define different regions of the quality score corresponding to quality classes and this allows the “evenly spaced” assumption to be tested. The figure above shows that the article quality levels are not evenly spaced! Interestingly, the way that they are unevenly spaced depends on how quality is measured. If the quality scale is designed to be accurate across all quality classes, then the quality levels are relatively even. However, if it is designed to be accurate for revisions or articles then more of the scale goes to lower levels of quality. Overall, the models agree that the difference between C-class and Start articles is smaller than that between the other levels.

Using the quality scores based on ordinal regression also improves accuracy. This mostly comes from calibrating the ordinal model to the appropriate unit of analysis. The ORES quality model is fit on a “balanced” dataset where each quality class is equally represented. This means that the ORES quality model has learned that each quality class is equally likely to occur in the data. Of course, in reality lower quality articles are far more common than high quality articles. There are over 3,000,000 Stubs on English Wikipedia but less than 8,000 Featured articles. As the table below shows, fitting ordinal models that know the true proportion of each quality class in a sample can improve accuracy compared to the ORES quality model.

Accuracy of quality prediction models depends on the unit of analysis. The greatest accuracy scores are highlighted. Models are more accurate when calibrated on the same unit of analysis on which they are evaluated. Compared to the MPQC, the ordinal quality models have better accuracy when revisions or articles are the unit of analysis. When the quality class is the unit of analysis, the ordinal quality model has slightly worse accuracy.

Even though my study has found evidence against the “evenly spaced” assumption, I also found that the quality scores based on it are strongly correlated with the scores from the ordinal model as shown in the figure below. The ‘𝜏’ symbols in the figure stand for the Kendall rank correlation coefficient while the ‘r’ symbols stand for the Pearson correlation coefficient. I used the Kendall correlation because it can capture nonlinear divergences between the measures and the Pearson coefficient is the familiar linear correlation. The “evenly spaced” scores aren’t totally in agreement with the scores from the ordinal models, but they are close enough that I doubt that prior research that used the “evenly spaced” scores to measure quality was mislead by this choice.

Correlations between quality measures show that the different approaches to measuring quality are quite similar. “Evenly spaced” uses the weighted sum of the ORES scores with handpicked coefficients [0,1,2,3,4,5]. Lower values of Kendall’s 𝜏, a nonparametric rank correlation statistic, compared to Pearson’s 𝑟 suggest nonlinear differences between the weighted sum and the other measures.

Conclusions

Measuring article quality in one continuous dimension is a valuable tool for studying the peer production of information goods because it provides granularity and is amenable to statistical analysis. Prior approaches extended ORES article quality prediction into a continuous measure under the “evenly spaced” assumption. I used ordinal regression models to transform the ORES predictions into a continuous measure of quality that is interpretable as a probability distribution over article quality levels, provides an account of its own uncertainty and does not assume that quality levels are “evenly spaced.” Calibrating the models to the chosen unit of analysis improves accuracy for research applications. I recommend that future work adopt this approach when article quality is an independent variable in a statistical analysis.

My paper also has a number of other tidbits about the uncertainty of different quality measures, the importance of feedback between measurement and knowledge in the scientific process and demonstrates model calibration.

A preprint of the paper is available here. The paper has been accepted to OpenSym 2021 and will be presented at the virtual conference on September 17th. A video of the presentation is available here. A dataverse repository containing code and data from the project is available here.

Posted on

Community Data Science Collective at ICA 2019 in Washington, DC

Jeremy Foote, Nate TeBlunthuis,  Wm Salt Hale, and Mako Hill will all be in Washington DC this week for the  International Communication Association’s 2019 annual meeting.

In particular, we be presenting a new paper from the group led by Sneha Narayan titled “All Talk: How Increasing Interpersonal Communication on Wikis May Not Enhance Productivity.” The talk will be on Monday, May 27 in a session from 9:30 to 10:45 in Washington Hilton LL, Holmead as part of a session organized by the ICA Computational Methods section on “Computational Approaches to Health Communication.”

Additionally, Nate is co-organizing a pre-conference at ICA on “Expanding Computational Communication: Towards a Pipeline for Graduate Students and Early Career Scholars” along with Josephine Lukito (UW Madison) and Frederic Hopp (UC Santa Barbara). The pre-conference will be held at American University on Friday May 24th. As part of that workshop, Nate and Jeremy will be giving a presentation on approaches to the study organizational communication that use computational methods.

We look forward to sharing our research and socializing with you at ICA! Please be in touch if you’re around and want to meet up!

Posted on

Revisiting the ‘Rise and Decline’

This graph shows the number of people contributing to Wikipedia over time:

The Rise and Decline of Wikipedia
The number of active Wikipedia contributors exploded, suddenly stalled, and then began gradually declining. (Figure taken from Halfaker et al. 2013)

The figure comes from “The Rise and Decline of an Open Collaboration System,” a well-known 2013 paper that argued that Wikipedia’s transition from rapid growth to slow decline in 2007 was driven by an increase in quality control systems. Although many people have treated the paper’s finding as representative of broader patterns in online communities, Wikipedia is a very unusual community in many respects. Do other online communities follow Wikipedia’s pattern of rise and decline? Does increased use of quality control systems coincide with community decline elsewhere?

In a paper I am presenting Thursday morning at  the Association for Computing Machinery (ACM) Conference on Human Factors in Computing Systems (CHI),  a group of us have replicated and extended the 2013 paper’s analysis in 769 other large wikis. We find that the dynamics observed in Wikipedia are a strikingly good description of the average Wikia wiki. They appear to reoccur again and again in many communities.

The original “Rise and Decline” paper (I’ll abbreviate it “RAD”) was written by Aaron Halfaker, R. Stuart Geiger, Jonathan T. Morgan, and John Riedl. They analyzed data from English Wikipedia and found that Wikipedia’s transition from rise to decline was accompanied by increasing rates of newcomer rejection as well as the growth of bots and algorithmic quality control tools. They also showed that newcomers whose contributions were rejected were less likely to continue editing and that community policies and norms became more difficult to change over time, especially for newer editors.

Our paper, just published in the CHI 2018 proceedings, replicates most of RAD’s analysis on a dataset of 769 of the  largest wikis from Wikia that were active between 2002 to 2010.  We find that RAD’s findings generalize to this large and diverse sample of communities.

I can walk you through some of the key findings. First, the growth trajectory of the average wiki in our sample is similar to that of English Wikipedia. As shown in the figure below, an initial period of growth stabilizes and leads to decline several years later.

Rise and Decline on Wikia
The average Wikia wikia also experience a period of growth followed by stabilization and decline (from TeBlunthuis, Shaw, and Hill 2018).

We also found that newcomers on Wikia wikis were reverted more and continued editing less. As on Wikipedia, the two processes were related. Similar to RAD, we also found that newer editors were more likely to have their contributions to the “project namespace” (where policy pages are located) undone as wikis got older. Indeed, the specific estimates from our statistical models are very similar to RAD’s for most of these findings!

There were some parts of the RAD analysis that we couldn’t reproduce in our context. For example, there are not enough bots or algorithmic editing tools in Wikia to support statistical claims about their effects on newcomers.

At the same time, we were able to do some things that the RAD authors could not.  Most importantly, our findings discount some Wikipedia-specific explanations for a rise and decline. For example, English Wikipedia’s decline coincided with the rise of Facebook, smartphones, and other social media platforms. In theory, any of these factors could have caused the decline. Because the wikis in our sample experienced rises and declines at similar points in their life-cycle but at different points in time, the rise and decline findings we report seem unlikely to be caused by underlying temporal trends.

The big communities we study seem to have consistent “life cycles” where stabilization and/or decay follows an initial period of growth. The fact that the same kinds of patterns happen on English Wikipedia and other online groups implies a more general set of social dynamics at work that we do not think existing research (including ours) explains in a satisfying way. What drives the rise and decline of communities more generally? Our findings make it clear that this is a big, important question that deserves more attention.

We hope you’ll read the paper and get in touch by commenting on this post or emailing me if you’d like to learn or talk more. The paper is available online and has been published under an open access license. If you really want to get into the weeds of the analysis, we will soon publish all the data and code necessary to reproduce our work in a repository on the Harvard Dataverse.

I will be presenting the project this week at CHI in Montréal on Thursday April 26 at 9am in room 517D.  For those of you not familiar with CHI, it is the top venue for Human-Computer Interaction. All CHI submissions go through double-blind peer review and the papers that make it into the proceedings are considered published (same as journal articles in most other scientific fields). Please feel free to cite our paper and send it around to your friends!

This blog post, and the open access paper that it describes, is a collaborative project with Aaron Shaw, and Benjamin Mako Hill. Financial support came from the US National Science Foundation (grants IIS-1617129,  IIS-1617468, and GRFP-2016220885 ), Northwestern University, the Center for Advanced Study in the Behavioral Sciences at Stanford University, and the University of Washington. This project was completed using the Hyak high performance computing cluster at the University of Washington.

Posted on

Community Data Science Collective at ICA 2017

A good chunk of the collective is heading to San Diego this week for the 2017 international communication association conference.

Here is a list of our ICA presentations, with links to the conference program which includes abstracts and other details:

In addition to papers, Aaron Shaw will also be chairing of the Critical Digital Labor and Algorithmic Studies session. Mon, May 29, 14:00 to 15:15, Hilton San Diego Bayfront, 2, Indigo 202A

We look forward to sharing research and socializing with you at ICA!

Posted on

Searching for competition on Change.org with LDA topic models

You may have heard of Change.org. It’s a popular online petitioning platform. You may have even noticed there can many online petitions about popular topics. For instance, it is easy to find dozens of petitions protesting the Lychee and Dog Meat Festival with varying levels of support.

Imagine you want to start an online petition. You might worry if your petition is very similar to other people’s petitions that already have signatures. These other petitions have a head start and will get all the attention. That said, if nobody has made any similar petitions, maybe that’s because the issue you are petitioning about doesn’t yet have a lot of popular support. You might also worry if your petition is unusual. Which of these two worries (making a duplicate petition and making a petition no one cares about) should concern you, dear petition creator? In my research, I set out to answer this question. The project is still in progress. I recently presented it as a poster at CSCW ’17.

Sociologists of organizational ecology considered similar questions about businesses and social movement organizations. They wanted to explain why organizations were more likely to die when an industry was young or old, but less likely to die in between. They argued that density, or the number of organizations in the population, was tied both to processes of legitimation and competition. There aren’t many firms in unproven industries because it’s not clear the industry will succeed, but when an industry is mature it becomes competitive. Everybody wants a piece of the pie, but you might not get enough pie to survive! This notion is called density dependence theory.

I think it is intuitive to apply this logic to online petitions and topics. If you make a petition about a low-density topic, chances for success should be lower because the petition is more likely to be unusual or illegitimate. However if you make a petition in a high-density topic, now you have to worry about competition with all the other petitions in the topic. You want your petition to be original, but not weird!

To collect data to test this theory, I downloaded a large set of petitions from Change.org, spam filtered them, and removed very short ones. Next I used LDA topic modeling to group petitions into topics. This makes it possible to assign petitions to points in a topic space. The more crowded this part of topic space, the denser the petition’s environment.

Finally, I used a regression model to predict petition signature counts. Since density dependence theory predicts that the relationship between density and signature count is shaped like an upside-down U, I included a quadratic term for density. The plot below shows that observed relationship between density in topic space and signature count is what the theory predicted. The darkness of the lines at the bottom of the plot show that most petitions are in less dense parts of topic space.  So you, dear petition creator, should worry about competition and legitimacy, but worry about legitimacy first!

I’m excited by this result because it shows interesting similarities between efforts to organize coordinated activism online and traditional organizations like firms. I’m planning to apply this method to other forms of online coordination like wikis and online communities.

This blog-post and the work it describes is a collaborative project between Nate TeBlunthuis, Benjamin Mako Hill and Aaron Shaw. We are still at work writing this project up as a research article. The work has been supported by the US National Science Foundation