In writing about experimentation, they report that Hal Varian, Google’s Chief Economist, estimates that Google runs “100-200 experiments on any given day”. This struck me as incredibly low! I would have guessed more like 10,000 or maybe more like 100,000.

The trick of course is how one individuates experiments. Say Google has an automatic procedure whereby each ad has a (small) random set of users who are prevented from seeing it and are shown the next best ad instead. Is this one giant experiment? Or one experiment for each ad?

This is a bit of a silly question.³

But when most people — even statisticians and scientists — think of an experiment in this context, they think of something like Google or Amazon making a particular button bigger. (Maybe somebody thought making that button bigger would improve a particular metric.) They likely don’t think of automatically generating an experiment for every button, such that a random sample see that particular button slightly bigger. It’s these latter kinds of procedures that lead to thinking about tens of thousands of experiments.

That’s the real deluge of experiments.

1. I don’t know that I would call much of it ‘innovation’. There is some outright innovation, but a lot of that is in the general strategies for using the data. There is much more gained in minor tweaking and optimization of products and services.
2. Perhaps they even overstate the power of simple experiments. For example, they do not mention the fact that many times the results these kinds of experiments often change over time, so that what you learned 2 months ago is no longer true.
3. Note that two single-factor experiments over the same population with independent random assignment can be regarded as a single experiment with two factors.

In fact, Maurits Kaptein and I use this term, coined by BJ Fogg, precisely because it sounds scary. We see the potential for quite negative consequences of persuasion profiling, so we try to alert our readers to this.¹

On the other hand, we also think that, not only is persuasion profiling sometimes beneficial, but there are cases where choosing not to adapt to individual differences in this way might itself be unethical.² If a company marketing a health intervention knows that there is substantial variety in how people respond to the strategies used in the intervention — such that while the intervention has positive effects on average, it has negative effects for some — it seems like they have two ethical options.

First, they can be honest about this in their marketing, reminding consumers that it doesn’t work for everyone or even trying to market it to people it is more likely to work for. Or they could make this interactive intervention adapt to individuals — by persuasion profiling.

Actually for the first option to really work, the company needs to at least model how these responses vary by observable and marketable-to characteristics (e.g., demographics). And it may be that this won’t be enough if there is too much heterogeneity: even within some demographic buckets, the intervention may have negative effects for a good number of would-be users. On the other hand, by implementing persuasion profiling, the intervention will help more people, and the company will be able to market it more widely — and more ethically.

A simplified example that is somewhat compelling to me at least, but certainly not airtight. In another post, I’ll describe how somewhat foreseeable, but unintended, consequences should also give one pause.

1. You might say we tried to build in a warning for anyone discussing or promoting this work.
2. We argue this in the paper we presented at Persuasive Technology 2010. The text below reprises some of what we said about our “Example 4″ in that paper.
   Kaptein, M. & Eckles, D. (2010). Selecting effective means to any end: Futures and ethics of persuasion profiling. Proceedings of Persuasive Technology 2010, Lecture Notes in Computer Science. Springer.

How useful are these traits for explanation, prediction, and adaptive interaction? I can’t address all of this here, but I want to sketch an argument for their irrelevance to adaptive interaction — and then offer a tentative rejoinder.

Interactive technologies can tailor their messages to the tastes and susceptibilities of the people interacting with and through them. It might seem that these traits should figure in the statistical models used to make these adaptive selections. After all, some of the possible messages fit for, e.g., coaching a person to meet their exercise goals are more likely to be effective for low NFC people than high NFC people, and vice versa. However, the standard questionnaire measures of NFC cannot often be obtained for most users — certainly not in commerce settings, and even people signing up for a mobile coaching service likely don’t want to answer pages of questions. On the other hand, some Internet and mobile services have other abundant data available about their users, which could perhaps be used to construct an alternative measure of these traits. The trait-based-adaptation recipe is:

1. obtain the questionnaire measure of the trait for a sample,
2. predict this measure with data available for many individuals (e.g., log data),
3. use this model to construct a measure for out-of-sample individuals.

This new measure could then be used to personalize the interactive experience based on this trait, such that if a version performs well (or poorly) for people with a particular score on the trait, then use (or don’t use) that version for people with similar scores.

But why involve the trait at all? Why not just personalize the interactive experience based on the responses of similar others? Since the new measure of the trait is just based on the available behavioral, demographic, and other logged data, one could simply predict responses based on those measure. Put in geometric terms, if the goal is to project the effects of different message onto available log data, why should one project the questionnaire measure of the trait onto the available log data and then project the effects onto this projection? This seems especially unappealing if one doesn’t fully trust the questionnaire measure to be accurate or one can’t be sure about which the set of all the traits that make a (substantial) difference.

I find this argument quite intuitively appealing, and it seems to resonate with others.¹ But I think there are some reasons the recipe above could still be appealing.

One way to think about this recipe is as dimensionality reduction guided by theory about psychological traits. Available log data can often be used to construct countless predictors (or “features”, as the machine learning people call them). So one can very quickly get into a situation where the effective number of parameters for a full model predicting the effects of different messages is very large and will make for poor predictions. Nothing — no, not penalized regression, not even a support vector machine — makes this problem go away. Instead, one has to rely on the domain knowledge of the person constructing the predictors (i.e., doing the “feature engineering”) to pick some good ones.

So the tentative rejoinder is this: established psychological traits might often make good dimensions to predict effects of different version of a message, intervention, or experience with. And they may “come with” suggestions about what kinds of log data might serve as measures of them. They would be expected to be reusable across settings. Thus, I think this recipe is nonetheless deserves serious attention.

1. I owe some clarity on this to some conversations with Mike Nowak, Maurits Kaptein, and others.

Here’s Natasha Singer at The New York Times on cascades of memes, idioms, and links through online social networks (e.g., Twitter):

If we keep seeing the same links and catchphrases ricocheting around our social networks, it might mean we are being exposed only to what we want to hear, says Damon Centola, an assistant professor of economic sociology at the Massachusetts Institute of Technology.

“You might say to yourself: ‘I am in a group where I am not getting any views other than the ones I agree with. I’m curious to know what else is out there,’” Professor Centola says.

Consider a new hashtag: diversity.

This is how Singer ends this article in which the central example is “icantdateyou” leading Egypt-related idioms as a trending topic on Twitter. The suggestion here, by Centola and Singer, is that people will notice they are getting a biased perspective of how many people agree with them and what topics people care about — and then will take action to get other perspectives.

Why am I skeptical?

First, I doubt that we really realize the extent to which media — and personalized social media in particular — bias their perception of the frequency of beliefs and events. Even though people know that fiction TV programs (e.g., cop shows) don’t aim to represent reality, heavy TV watchers (on average) substantially overestimate the percent of adult men employed in law enforcement.¹ That is, the processes that produce the “friendly world syndrome” function without conscious awareness and, perhaps, even despite it. So people can’t consciously choose to seek out diverse perspectives if they don’t know they are increasingly missing them.

Second, I doubt that people actually want diversity of perspectives all that much. Even if I realize divergent views are missing from my media experience, why would I seek them out? This might be desirable for some people (but not all), and even for those, the desire to encounter people who radically disagree has its limits.

Similar ideas pop up in a NPR All Things Considered segment by Laura Sydell. This short piece (audio, transcript) is part of NPR’s “Cultural Fragmentation” series. The segment begins with the worry that offline bubbles are replicated online and quotes me describing how attempts to filter for personal relevance also heighten the bias towards agreement in personalized media.

But much of the piece has actually focuses on how one person — Kyra Gaunt, a professor and musician — is using Twitter to connect and converse with new and different people. Gaunt describes her experience on Twitter as featuring debate, engagement, and “learning about black people even if you’ve never seen one before”. Sydell’s commentary identifies the public nature of Twitter as an important factor in facilitating experiencing diverse perspectives:

But, even though there is a lot of conversation going on among African Americans on Twitter, Professor Gaunt says it’s very different from the closed nature of Facebook because tweets are public.

I think this is true to some degree: much of the content produced by Facebook users is indeed public, but Facebook does not make it as easily searchable or discoverable (e.g., through trending topics). But more importantly, Facebook and Twitter differ in their affordances for conversation. Facebook ties responses to the original post, which means both that the original poster controls who can reply and that everyone who replies is part of the same conversation. Twitter supports replies through the @reply mechanism, so that anyone can reply but the conversation is fragmented, as repliers and consumers often do not see all replies. So, as I’ve described, even if you follow a few people you disagree with on Twitter, you’ll most likely see replies from the other people you follow, who — more often than not — you agree with.

Gaunt’s experience with Twitter is certainly not typical. She has over 3,300 followers and follows over 2,400, so many of her posts will generate replies from people she doesn’t know well but whose replies will appear in her main feed. And — if she looks beyond her main feed to the @Mentions page — she will see the replies from even those she does not follow herself. On the other hand, her followers will likely only see her posts and replies from others they follow.²

Nonetheless, Gaunt’s case is worth considering further, as does Sydell:

SYDELL: Gaunt says she’s made new friends through Twitter.

GAUNT: I’m meeting strangers. I met with two people I had engaged with through Twitter in the past 10 days who I’d never met in real time, in what we say in IRL, in real life. And I met them, and I felt like this is my tribe.

SYDELL: And Gaunt says they weren’t black. But the key word for some observers is tribe. Although there are people like Gaunt who are using social media to reach out, some observers are concerned that she is the exception to the rule, that most of us will be content to stay within our race, class, ethnicity, family or political party.

So Professor Gaunt is likely making connections with people she would not have otherwise. But — it is at least tempting to conclude from “this is my tribe” — they are not people with radically different beliefs and values, even if they have arrived at those beliefs and values from a membership in a different race or class.

1. Gerbner, G., Gross, L., Morgan, M., & Signorielli, N. (1980). The “Mainstreaming” of America: Violence Profile No. 11. Journal of Communication, 30(3), 10-29.
2. One nice feature in “new Twitter” — the recently refresh of the Twitter user interface — is that clicking on a tweet will show some of the replies to it in the right column. This may offer an easier way for followers to discover diverse replies to the people they follow. But it is also not particularly usable, as it is often difficult to even trace what a reply is a reply to.

Aardvark employees would get the questions from beta test users and route them to users who were online and would have the answer to the question. This was done to test out the concept before the company spent the time and money to build it, said Damon Horowitz, co-founder of Aardvark, who spoke at Startup Lessons Learned, a conference in San Francisco on Friday.

“If people like this in super crappy form, then this is worth building, because they’ll like it even more,” Horowitz said of their initial idea.

At the same time it was testing a “fake” product powered by humans, the company started building the automated product to replace humans. While it used humans “behind the curtain,” it gained the benefit of learning from all the questions, including how to route the questions and the entire process with users.

This is a really good idea, as I’ve argued before on this blog and in a chapter for developers of mobile health interventions. What better way to (a) learn about how people will use and experience your service and (b) get training data for your machine learning system than to have humans-in-the-loop run the service?

My friend Chris Streeter wondered whether this was all done by Aardvark employees or whether workers on Amazon Mechanical Turk may have also been involved, especially in identifying the expertise of the early users of the service so that the employees could route the questions to the right place. I think this highlights how different parts of a service can draw on human and non-human intelligence in a variety of ways — via a micro-labor market, using skilled employees who will gain hands-on experience with customers, etc.

I also wonder what UIs the humans-in-the-loop used to accomplish this. It’d be great to get a peak. I’d expect that these were certainly rough around the edges, as was the Aardvark customer-facing UI.

Aardvark does a good job of being a quite sociable agent (e.g., when using it via instant messaging) that also gets out of the way of the human–human interaction between question askers and answers. I wonder how the language used by humans to coordinate and hand-off questions may have played into creating a positive para-social interaction with vark.

Instead, I want to relate this discussion to my continued interest in how activity streams and other information consumption interfaces affect their users’ beliefs and behaviors through the availability heuristic. In response to some comments on his first post, Stutzman argues that people overestimate the degree to which content once public on the Internet is public forever:

So why is it that we all assume that the content we share publicly will be around forever?  I think this is a classic case of selection on the dependent variable.  When we Google ourselves, we are confronted with what’s there as opposed to what’s not there.  The stuff that goes away gets forgotten, and we concentrate on things that we see or remember (like a persistent page about us that we don’t like).  In reality, our online identities decay, decay being a stochastic process.  The internet is actually quite bad at remembering.

This unconsidered “selection on the dependent variable” is one way of thinking about some cases of how the availability heuristic (and use of ease-of-retrievel information more generally). But I actually think the latter is more general and more useful for describing the psychological processes involved. For example, it highlights both that there are many occurrences or interventions can can influence which cases are available to mind and that even if people have thought about cases where their content disappeared at some point, this may not be easily retrieved when making particular privacy decisions or offering opinions on others’ actions.

Stutzman’s example is but one way that the combination of the availability heuristic and existing Internet services combine to affect privacy decisions. For example, consider how activity streams like Facebook News Feed influence how people perceive their audience. News Feed shows items drawn from an individual’s friends’ activities, and they often have some reciprocal access. However, the items in the activity stream are likely unrepresentative of this potential and likely audience. “Lurkers” — people who consume but do not produce — are not as available to mind, and prolific producers are too available to mind for how often they are in the actual audience for some new shared content. This can, for example, lead to making self-disclosures that are not appropriate for the actual audience.

1. This might not be the case, see Michael Zimmer and this New York Times article.
2. Why don’t people do this in the first place? Many may not be aware of the feature, but even if they are, there are reasons not to use it. For example, it makes any participation in topical conversations (e.g., around a hashtag) difficult or impossible.
3. Or at least this control would have to be via Twitter, likely before archiving: “We asked them [Twitter] to deal with the users; the library doesn’t want to mediate that.”

“Supertaskers did a phenomenal job of performing several different tasks at once,” Watson says. “We’d all like to think we could do the same, but the odds are overwhelmingly against it.” (Wired News & Science News)

The researchers, Watson and Strayer, argue that they have good evidence for the existence of this individual variation. One can find many media reports of this “discovery” of “supertaskers” (e.g., Psychology Today). I do not think this conclusion is well justified.

First, let’s consider the methods used in this research. 100 college students each completed driving tasks and an auditory task on a mobile phone — separately and in combination — over a single 1.5 hour session. The auditory task is designed to measure differences in executive attention by requiring participants do hold past items in memory while completing math tasks. The researchers identified “supertaskers” as those participants who met the following “stringent” requirements: they were both (a) in the top 25% of participants in performance in the single-task portions and (b) and not different in their dual-task performance on at least three of the four measures by more than the standard error. Since two of the four measures are associated with each of the two tasks (driving: brake reaction time, following distance; mobile phone task: memory performance, math performance), this requires that ”supertaskers” do as well on both measures of either the driving or mobile phone task and one measure of the other task.

There may be many issues with the validity of the inference in this work. I want to focus on one in particular: the inference from the observation of differences between participants’ performance in a single 1.5 hour session to the conclusion that there are stable, “trait” differences among participants, such that some are “supertaskers”. This conclusion is simply not justified. To illustrate this, let’s consider how the methods of this study differ from those usually (and reasonably) used by psychologists to reach such conclusions.

Psychologists often study individual differences using the following approach. First, identify some plausible trait of individuals. Second, construct a questionnaire or other (perhaps behavioral) test that measures that trait. Third, demonstrate that this test has high reliability — that is, that the differences between people are much larger than the differences between the same person taking the test at different times. Fourth, then use this test to measure the trait and see if it predicts differences in some experiment. A key point here is that in order to conclude that the test measures a stable individual difference (i.e., a trait) researchers need to establish high test-retest reliability; otherwise, the test might just be measuring differences in temporary mood.

Returning to Watson and Strayer’s research, it is easy to see the problem: we have no idea whether the variation observed should be attributed to stable individual differences (i.e., being a “supertasker”) or to unstable differences. That is, if we brought those same “supertasker” participants back into the lab and they did another session, would they still exhibit the same lack of performance difference between the single- and dual-task conditions? This research gives us no reason that expect that they would.

Watson and Strayer do some additional analysis with the aim of ruling out their observations being a fluke. One might think this addresses my criticism, but it does not. They

performed a Monte Carlo simulation in which randomly selected single-dual task pairs of variables from the existing data set were obtained for each of the 4 dependent measures and then subjected to the same algorithm that was used to classify the supertaskers.

That is, they broke apart the single-task and dual-task data for each participant and created new simulated participants by randomly sampling pairs single- and dual-task data. They found that on this analysis there would be only 1/15th of the observed ”supertaskers”. This is a good analysis to do. However, this just demonstrates that being labeled a “supertasker” is likely caused by the single- and dual-task data being generated by the same person in the same session. This stills leaves it quite open (and more plausible to me) that participants’ were in varying states for the session and this explains their (temporary) “supertasking”. It also allows that this greater frequency of “supertaskers” is due to participants who do well in whatever task they are given first being more likely to do well in subsequent tasks.

My aim in this post is to suggest some challenges that this kind of approach has to face. Part of my interest in this is that I’m quite sympathetic to identifying stable, observed differences in behavior and then “working backwards” to characterizing the traits that explain these downstream differences. This  exactly the approach that Maurits Kaptein and I are taking in our work on persuasion profiling: we observe how individuals respond to the use of different influence strategies and use this to (a) construct a “persuasion profile” for that individual and (b) characterize how much variation in the effects of these strategies there is in the population.

However, a critical step in this process is ruling out the alternative explanation that the observed differences are primarily due to differences in, e.g., mood, rather than stable individual differences. One way to do this is to observe the behavior in multiple sessions and multiple contexts. Another way to rule out this alternative explanation is if you observe a complex pattern of behavioral differences that previous work suggests could not be the result of temporary, unstable differences — or at least is more easily explained by previous theories about the relevant traits. That is, I’m enthusiastic about identifying stable, observed differences in behavior, but I don’t want to see researchers abandon the careful methods that have been used in the past to make the case for a new individual difference.

Watson, Strayer, and colleagues have apparently begun doing work that could be used to show the stability of the observed differences. The discussion section of their paper refers to some additional unpublished research in which they invited their “supertaskers” from this study and another study back into the lab and had them do some similar tasks measuring executive attention (but not driving) while in an fMRI machine. They report greater “coherence” in their performance in this second study and the previous study than control participants and better performance for “supertaskers” on dual-N-back tasks. But this is short of showing high test-retest reliability.

Since little is said about this work, I hesitate to conclude anything from it or criticize it. I’ve contacted the authors with the hope of learning more. My current sense is that Watson and Strayer’s entire case for “supertaskers” hinges on research of this kind.

References

My thinking on persuasion profiling is very much in BJ Fogg’s footsteps, since he has been talking about persuasion profiling in courses, lab meetings, and personal discussions since 2004 or earlier.

Just yesterday, I came across this transcript of BJ’s presentation for an FTC hearing in 2006. I was struck at how much it anticipates some of what Maurits and I have written recently (more on this later). I’m sure I watched the draft video of the presentation back then and it’s influenced me, even if I forgot some of the details.

Here is the relevant excerpt from BJ’s comments for the FTC:

Persuasion profiling means that each one of us has a different set of persuasion strategies that affect us. Just like we like different types of food or are vulnerable to giving in to different types of food on a diet, we are vulnerable to different types of persuasion strategies.

On the food example, I love old-fashioned popcorn, and if I go to a party and somebody has old-fashioned popcorn, I will probably break down and eat it. On the persuasion side of things, I know I’m vulnerable to trying new things, to challenges and to anything that gets measured. If that’s proposed to me, I’m going to be vulnerable and I’m going to give it a shot.

Whenever we go to a Web site and use an interactive system, it is likely they will be capturing what persuasion strategies work on us and will be using those when we use the service again. The mapping out of what makes me tick, what motivates me can also be bought or sold, just like a credit report.

So imagine I’m going in to buy a new car and the person selling me the car downloads my credit report but also buys my persuasion profile. I may or may not know about this. Imagine if persuasion profiles are available on political campaigns so that when I visit a Web site, the system knows it is B.J. Fogg, and it changes [its] approach based on my vulnerabilities when it comes to persuasion.

Persuasive technology will touch our lives anywhere that we access digital products or services, in the car, in our living room, on the Web, through our mobile phones and so on. Persuasive technology will be all around us, and unlike other media types, where you have 30-second commercial or a magazine ad, you have genres you can understand, when it comes to computer-based persuasion, it is so flexible that it won’t have genre boundaries. It will come to us in the ordinary course of our lives, as we are working on a Web site, as we are editing a document, as we are driving a car. There won’t be clear markers about when you are being persuaded and when you are not.

This last paragraph is about the “genrelessness” of many persuasive technologies. This isn’t directly on the topic of persuasion profiling, but I see it as critically relevant. Persuasion profiling is likely to be most effective when invisible and undisclosed to users. From this and the lack of genre-based flags for persuasive technology it follows that we will frequently be “persuasion profiled” without knowing it.

In finding relevant research, once one has found something interesting, it can be really useful to do “reverse citation” searches.

Google Scholar is often my first stop when finding research literature (and for general search), and it has this feature — just click “Cited by 394″. But it is not very useful when your starting point is highly cited. What I often want to do is to do a keyword search of the papers that cite my highly-cited starting point.

While there is no GUI for this search within these resultsin Google Scholar, you can actually do it by hacking the URL. Just add the keyword query to the URL.

This is the URL one gets for all resources Google has as citing Allport’s “Attitudes” (1935):

http://scholar.google.com/scholar?cites=9150707851480450787&hl=en

And this URL searches within those for “indispensable concept”:

http://scholar.google.com/scholar?hl=en&cites=9150707851480450787&q=indispensable+concept

In this particular case, this gives us many examples of authors citing Allport’s comment that the attitude is the most distinctive and indispensable concept in social psychology. This example highlights that this can even just help get more useful “snippets” in the search results, even if it doesn’t narrow down the results much.

I find this useful in many cases. Maybe you will also.

Chess Wars was being rejected after the six week wait [because] the bubbles in its chat rooms are too shiny, and Apple has trademarked that bubbly design. [...] The representative said Stump needed to make the bubbles “less shiny” and also helpfully suggested that he make the bubbles square, just to be sure.

One thing that is quite striking in this situation is that it is at odds with Apple’s long history of strongly encouraging third-party developers to follow many UI guidelines — guidelines that when followed make third-party apps blend in like they’re native.¹

It’s important to not read too much into this (especially since we don’t know what Apple’s more considered policy on this will end up being), but it is interesting to think about how responsibility gets spread around among mobile applications, services, and devices — and how this may be different than existing models on the desktop.My sense is that experienced desktop computer users understand at least the most important ways sources of their good and bad experiences are distinguished. For example, “locomotion” is a central metaphor in using the Web, as opposed to the conversation and manipulation metaphors of the command line / natural language interfaces and WIMP: we “go to” a site (see this interview with Terry Winograd, full .mov here). The locomotion metaphor helps people distinguish what my computer is contributing and what some distant, third-party “site” is contributing.

This is complex even on the Web, but many of these genre rules are currently being all mixed up. Google has Gmail running in your browser but on your computer. Cameraphones are recognizing objects you point them at — some by analyzing the image on the device and some by sending the device to a server to be analyzed.

This issue is sometimes identified by academics as one of source orientation and source equivocality. Though there has been some great research in this area, there is a lot we don’t know and the field is in flux: people’s beliefs about systems are changing and the important technologies and genres are still emerging.

If there’s one important place to start thinking about the craziness of the current situation of ubiquitous source equivocality is “Personalization of mass media and the growth of pseudo-community” (1987) by James Beniger that predates much of the tech at issue.

1. I was led to think this by a commenter on TechCrunch, Dan Grossman, pointing out this long history.