Skip to content

{ Tag Archives } BALANCE

Using Mechanical Turk for experiments

In my upcoming CHI paper, “Presenting Diverse Political Opinions: How and How Much,” we used Amazon’s Mechanical Turk (AMT) to recruit subjects and to administer the study. I’ll talk a bit more about the research questions and results in a future post, but I’ve had enough questions about using Mechanical Turk that I think a blog post may be helpful.

In this study, Paul Resnick and I explored individuals’ preferences for diversity of political opinion in online news aggregators and evaluated whether some very basic presentation techniques might affect satisfaction with the range of opinions represented in a collection of articles.

To address these questions, we needed subjects with known political preferences, from the United States, and with at least some very basic political knowledge, and we wanted to collect some demographic information about each subject. Each approved subject was then assigned to either a manipulation check group or to the experimental group. Subjects in the manipulation check group viewed individual articles and indicated their agreement or disagreement with each; subjects in the experimental group viewed entire collections and answered questions about the collection. The subjects in the experimental group were also assigned to a particular treatment (how the list would appear to them). Once approved, subjects could view a list up to once per day.

Screening. To screen subjects, we used a Qualification test in AMT. When unqualified subjects viewed at task (HIT – human intelligence task, in mTurk parlance), they were informed that they needed to complete a qualification. The qualification test asked subjects two questions about their political preferences, three multiple choice questions about US politics, and a number of demographic questions. Responses were automatically downloaded and evaluated to complete screening and assignment.

To limit our subjects to US residents, we also used the automatic locale qualification.

Assignment. We handled subject assignment in two ways. To distinguish between the treatment group and the manipulation check group, we created to additional qualifications that were automatically assigned; an approved subject would be granted only one of these qualifications, and could thus could only complete the associated task type.

Tasks (HITs). The task implementation was straightforward. We hosted tasks on our own server using the external question HIT type. When a subject loaded a task, AMT passed us the subject’s worker ID. We verified that the subject was qualified for the task and loaded the appropriate presentation for that subject. Each day, we uploaded one task of each type, with many assignments; assignments are the number of turkers that can complete each task.

Because we needed real-time access to the manipulation check data, the responses to this task were stored in our own database after a subject submitted the form; the subject could then return to AMT. This was not necessary for the experimental data, and so the responses were sent directly to AMT for later retrieval.

Quality control. Careless clicking or hurrying through the task is a potential problem on mTurk. Using multiple raters does not work when asking subjects about their opinions. Kittur, Chi, and Suh recommend asking Turkers verifiable questions as a way to deal with the problem1. We did not, however, ask verifiable questions about any of the articles or the list, because that might have changed how turkers read the list and responded to our other questions. Instead, we randomly repeated a demographic question from the qualification test. 5 subjects changed their answer substantially (e.g. aging more than one year or in reverse or shifting on either of the political spectrum questions by 2 points or more). Though there are many possible explanations for these shifts – such as shared accounts within a household, careless clicking, easily shifting political opinions, deliberate deception, or lack of effort – all of these explanations are not desirable for study subjects, and so they were excluded. We also examined how long each subject took to complete tasks (looking for implausibly fast responses); this did not lead to the exclusion of any additional subjects or responses.

Some reflection. We had to pay turkers a bit more than we expected (~$12/hr) and we recruited fewer subjects than we anticipated. The unpaid qualification task may be a bit of a barrier, especially because potential subjects could only complete one of our paid tasks per day (and only one was listed at a time). Instead, we might have implemented the qualification as a paid task, but that might result in paying for subjects who would never return to complete an actual task.

Further resources

1. Kittur, A., Chi, E. H., and Suh, B. (2009). “Crowdsourcing User Studies With Mechanical Turk,” Proc. CHI 2009: 453-456. (ACM | PARC)
2. Mason, W. and Watts, D. J. (2009). “Financial incentives and the ‘performance of crowds,’” SIGKDD Workshop on Human Computation: 77-85. (ACM | Yahoo)

This study is part of the BALANCE project and was funded by NSF award #IIS-0916099.

updated viz of political blogs’ link similarity

I’ve been meaning to post a simple update to my previous visualization of political blogs’ link similarities. In the previous post, I used GEM for layout, which was not, in hindsight, the best choice.

In the visualization in this post, the edges between blogs (the nodes, colorized as liberal, independent, and conservative) are weighted as the Jaccard similarity between any two blogs. The visualization is then laid out in GUESS using multidimensional scaling (MDS) based on the Jaccard similarities.

Sidelines at ICWSM

Last week I presented our first Sidelines paper (with Daniel Zhou and Paul Resnick) at ICWSM in San Jose. Slides (hosted on slideshare) are embedded below, or you can watch a video of most of the talk on VideoLectures.

Opinion and topic diversity in the output sets can provide individual and societal benefits. If news aggregators relying on votes and links to select and subsets of the large quantity of news and opinion items generated each day simply select the most popular items may not yield as much diversity as is present in the overall pool of votes and links.

To help measure how well any given approach does at achieving these goals, we developed three diversity metrics that address different dimensions of diversity: inclusion/exclusion, nonalienation, and proportional representation (based on KL divergence).

To increase diversity in result sets chosen based on user votes (or things like votes), we developed the sidelines algorithm. This algorithm temporarily suppresses a voter’s preferences after a preferred item has been selected. In comparison to collections of the most popular items, from user votes on and links from a panel of political blogs, the Sidelines algorithm increased inclusion while decreasing alienation. For the blog links, a set with known political preferences, we also found that Sidelines improved proportional representation.

Our approach differs and is complementary to work that selects for diversity or identifies bias based on classifying content (e.g. Park et al, NewsCube; ) or by classifying referring blogs or voters (e.g. Gamon et al, BLEWS). While Sidelines requires votes (or something like votes), it doesn’t require any information about content, voters, or long term voting histories. This is particularly useful for emerging topics and opinion groups, as well as for non-textual items.

visualizing political blogs’ linking

There are a number of visualizations of political bloggers’ linking behavior, notably Adamic and Glance’s 2005 work that found political bloggers of one bias tend to link to others of the same bias. Also check out Linkfluence’s Presidential Watch 08 map, which indicates similar behavior.

These visualizations are based on graphs of when one blog links to another. I was curious to what extent this two-community behavior occurs if you include all of the links from these blogs (such as links to news items, etc). Since I have link data for about 500 blogs from the news aggregator work, it was straightforward to visualize a projection of the bipartite blog->item graph. To classify each blog as liberal, conservative, or independent, I used a combination of the coding from Presidential Watch, Wonkosphere, and my own reading.

Projection of links from political blogs to items (Oct - Nov 2008)

Projection of links from political blogs to items (Oct - Nov 2008). Layout using GEM algorithm in GUESS.

The visualization shows blogs as nodes. Edges represent shared links (at least 6 items must be shared before drawing an edge) and are sized based on their weight. Blue edges run between liberal blogs, red edges between conservative blogs, maroon between conservative and independent, violet blue between liberal and independent, purple between independent blogs, and orange between liberal and conservative blogs. Nodes are sized as a log of their total degree. This visualization is formatted to appear similar to the Adamic and Glance graph, though there are some important differences, principally because this graph is undirected and because I have included independent blogs in the sample.

This is just a quick look, but we can see that the overall linking behavior still produces two fairly distinct communities, though a bit more connected than just the graph of blog to blog links. It’d be fun to remove the linked blog posts from this data (leaving mostly linked news items) to see if that changes the picture much. Are some media sources setting the agenda for bloggers of both parties, or are the conservative bloggers reading and reacting to one set news items and liberal bloggers reading and reacting to another? I.e., is the homophily primarily in links to opinion articles, or does it also extend to the linked news items?

I’m out of time at this point in the semester, though, so that will have to wait.

bias mining in political bloggers’ link patterns

I was pretty excited by the work that Andy Baio and Joshua Schachter did to identify and show the political leanings in the link behavior of blogs that are monitored by Memeorandum. They used singular value decomposition [1] on an adjacency matrix between sources and items based on link data from 360 snapshots of Memeorandum’s front page.

For the political news aggregator project, we’ve been gathering link data from about 500 blogs. Our list of sources is less than half of theirs (I only include blogs that make full posts available in their feeds), but we do have full link data rather than snapshots, so I was curious if we would get similar results.

The first 10 columns of two different U matrices are below. They are both based on link data from 3 October to 7 November; the first includes items that had an in-degree of at least 4 (5934 items), the second includes items with an in-degree of at least 3 (9722 items). In the first, the second column (v2) seems to correspond fairly well to the political leaning of the blog; in the second, the second column (v3) is better.

I’ll be the first to say that I haven’t had much time look at these results in any detail, and, as some of the commenters on Andy’s post noted, there are probably better approaches for identifying bias than SVD. If you’d like to play too, you can download a csv file with the sources and all links with an in-degree >= 2 (21517 items, 481 sources). Each row consists of the source title, source url, and then a list of the items the source linked to from 3 October to 7 November. Some sources were added part way though this window, and I didn’t collect link data from before they were added.

[1] One of the more helpful singular value decomposition tutorials I found was written by Kirk Baker and is available in PDF.

US political news and opinion aggregation

Working with Paul Resnick and Xiaodan Zhou, I’ve started a project to build political news aggregators that better reflect diversity and represent their users, even when there is an unknown political bias in the inputs. We’ll have more on this to say later, but for now we’re making available a Google gadget based on a prototype aggregator’s results.

The list of links is generated from link data from about 500 blogs and refreshed every 30 minutes. Some of the results will be news stories, some will be op-ed columns from major media services, others will be blog posts, and there are also some other assorted links.

At this early point in our work, the results tend to be more politically diverse than an aggregator such as Digg, but suffer from problems with redundancy (we aren’t clustering links about the same story yet). As our results get better, the set of links the gadget shows should improve.

Update 15 December: I twittered last week that I’ve added bias highlighting to the widget, but I should expand a bit on that here.

Inspired by Baio and Schachter’s coloring of political bias on Memeorandum, I’ve added a similar feature to the news aggregator widget. Links are colored according the average bias of the blogs linking to them. This is not always a good predictor of the item’s bias or whether it better supports a liberal or conservative view. Sometimes a conservative blogger writes a post to which more liberal bloggers than conservative bloggers, and in that case, the link will be colored blue.

If you don’t like the highlighting, you can turn it off in the settings.