Selecting effective means to any end
How are psychographic personalization and persuasion profiling different from more familiar forms of personalization and recommendation systems? A big difference is that they focus on selecting the “how” or the “means” of inducing you to an action — rather than selecting the “what” or the “ends”. Given the recent interest in this kind of personalization, I wanted to highlight some excerpts from something Maurits Kaptein and I wrote in 2010.1
This post excerpts our 2010 article, a version of which was published as:
Kaptein, M., & Eckles, D. (2010). Selecting effective means to any end: Futures and ethics of persuasion profiling. In International Conference on Persuasive Technology (pp. 82-93). Springer Lecture Notes in Computer Science.
For more on this topic, see these papers.
We distinguish between those adaptive persuasive technologies that adapt the particular ends they try to bring about and those that adapt their means to some end.
First, there are systems that use models of individual users to select particular ends that are instantiations of more general target behaviors. If the more general target behavior is book buying, then such a system may select which specific books to present.
Second, adaptive persuasive technologies that change their means adapt the persuasive strategy that is used — independent of the end goal. One could offer the same book and for some people show the message that the book is recommended by experts, while for others emphasizing that the book is almost out of stock. Both messages be may true, but the effect of each differs between users.
Example 2. Ends adaptation in recommender systems
Pandora is a popular music service that tries to engage music listeners and persuade them into spending more time on the site and, ultimately, subscribe. For both goals it is beneficial for Pandora if users enjoy the music that is presented to them by achieving a match between the music offering to individual, potentially latent music preferences. In doing so, Pandora adaptively selects the end — the actual song that is listened to and that could be purchased, rather than the means — the reasons presented for the selection of one specific song.
The distinction between end-adaptive persuasive technologies and means-adaptive persuasive technologies is important to discuss since adaptation in the latter case could be domain independent. In end adaptation, we can expect that little of the knowledge of the user that is gained by the system can be used in other domains (e.g. book preferences are likely minimally related to optimally specifying goals in a mobile exercise coach). Means adaptation is potentially quite the opposite. If an agent expects that a person is more responsive to authority claims than to other influence strategies in one domain, it may well be that authority claims are also more effective for that user than other strategies in a different domain. While we focus on novel means-adaptive systems, it is actually quite common for human influence agents adaptively select their means.
Influence Strategies and Implementations
Means-adaptive systems select different means by which to bring about some attitude or behavior change. The distinction between adapting means and ends is an abstract and heuristic one, so it will be helpful to describe one particular way to think about means in persuasive technologies. One way to individuate means of attitude and behavior change is to identify distinct influence strategies, each of which can have many implementations. Investigators studying persuasion and compliance-gaining have varied in how they individuate influence strategies: Cialdini [5] elaborates on six strategies at length, Fogg [8] describes 40 strategies under a more general definition of persuasion, and others have listed over 100 [16].
Despite this variation in their individuation, influence strategies are a useful level of analysis that helps to group and distinguish specific influence tactics. In the context of means adaptation, human and computer persuaders can select influence strategies they expect to be more effective that other influence strategies. In particular, the effectiveness of a strategy can vary with attitude and behavior change goals. Different influence strategies are most effective in different stages of the attitude to behavior continuum [1]. These range from use of heuristics in the attitude stage to use of conditioning when a behavioral change has been established and needs to be maintained [11]. Fogg [10] further illustrates this complexity and the importance of considering variation in target behaviors by presenting a two-dimensional matrix of 35 classes behavior change that vary by (1) the schedule of change (e.g., one time, on cue) and (2) the type of change (e.g., perform new behavior vs. familiar behavior). So even for persuasive technologies that do not adapt to individuals, selecting an influence strategy — the means — is important. We additionally contend that influence strategies are also a useful way to represent individual differences [9] — differences which may be large enough that strategies that are effective on average have negative effects for some people.
Example 4. Backfiring of influence strategies
John just subscribed to a digital workout coaching service. This system measures his activity using an accelerometer and provides John feedback through a Web site. This feedback is accompanied by recommendations from a general practitioner to modify his workout regime. John has all through his life been known as authority averse and dislikes the top-down recommendation style used. After three weeks using the service, John’s exercise levels have decreased.
Persuasion Profiles
When systems represent individual differences as variation in responses to influence strategies — and adapt to these differences, they are engaging in persuasion profiling. Persuasion profiles are thus collections of expected effects of different influence strategies for a specific individual. Hence, an individual’s persuasion profile indicates which influence strategies — one way of individuating means of attitude and behavior change — are expected to be most effective.
Persuasion profiles can be based on demographic, personality, and behavioral data. Relying primarily on behavioral data has recently become a realistic option for interactive technologies, since vast amounts of data about individuals’ behavior in response to attempts at persuasion are currently collected. These data describe how people have responded to presentations of certain products (e.g. e-commerce) or have complied to requests by persuasive technologies (e.g. the DirectLife Activity Monitor [12]).
Existing systems record responses to particular messages — implementations of one or more influence strategies — to aid profiling. For example, Rapleaf uses responses by a users’ friends to particular advertisements to select the message to present to that user [2]. If influence attempts are identified as being implementations of particular strategies, then such systems can “borrow strength” in predicting responses to other implementations of the same strategy or related strategies. Many of these scenarios also involve the collection of personally identifiable information, so persuasion profiles can be associated with individuals across different sessions and services.
Consequences of Means Adaptation
In the remainder of this paper we will focus on the implications of the usage of persuasion profiles in means-adaptive persuasive systems. There are two properties of these systems which make this discussion important:
1. End-independence: Contrary to profiles used by end-adaptive persuasive sys- tems the knowledge gained about people in means-adaptive systems can be used independent from the end goal. Hence, persuasion profiles can be used independent of context and can be exchanged between systems.
2. Undisclosed: While the adaptation in end-adaptive persuasive systems is often most effective when disclosed to the user, this is not necessarily the case in means-adaptive persuasive systems powered by persuasion profiles. Selecting a different influence strategy is likely less salient than changing a target behavior and thus will often not be noticed by users.
Although through the previous examples and the discussion of adaptive persuasive systems these two notions have already been hinted upon, we feel it is important to examine each in more detail.
End-Independence
Means-adaptive persuasive technologies are distinctive in their end-independence: a persuasion profile created in one context can be applied to bringing about other ends in that same context or to behavior or attitude change in a quite different context. This feature of persuasion profiling is best illustrated by contrast with end adaptation.
Any adaptation that selects the particular end (or goal) of a persuasive attempt is inherently context-specific. Though there may be associations between individual differences across context (e.g., between book preferences and political attitudes) these associations are themselves specific to pairs of contexts. On the other hand, persuasion profiles are designed and expected to be independent of particular ends and contexts. For example, we propose that a person’s tendency to comply more to appeals by experts than to those by friends is present both when looking at compliance to a medical regime as well as purchase decisions.
It is important to clarify exactly what is required for end-independence to obtain. If we say that a persuasion profile is end-independent than this does not imply that the effectiveness of influence strategies is constant across all contexts. Consistent with the results reviewed in section 3, we acknowledge that influence strategy effectiveness depends on, e.g., the type of behavior change. That is, we expect that the most effective influence strategy for a system to employ, even given the user’s persuasion profile, would depend on both context and target behavior. Instead, end-independence requires that the difference between the average effect of a strategy for the population and the effect of that strategy for a specific individual is relatively consistent across contexts and ends.
Implications of end-independence.
From end-independence, it follows that persuasion profiles could potentially be created by, and shared with, a number of systems that use and modify these profiles. For example, the profile constructed from observing a user’s online shopping behavior can be of use in increasing compliance in saving energy. Behavioral measures in latter two contexts can contribute to refining the existing profile.2
Not only could persuasion profiles be used across contexts within a single organization, but there is the option of exchanging the persuasion profiles between corporations, governments, other institutions, and individuals. A market for persuasion profiles could develop [9], as currently exists for other data about consumers. Even if a system that implements persuasion profiling does so ethically, once constructed the profiles can be used for ends not anticipated by its designers.
Persuasion profiles are another kind of information about individuals collected by corporations that individuals may or have effective access to. This raises issues of data ownership. Do individuals have access to their complete persuasion profiles or other indicators of the contents of the profiles? Are individuals compensated for this valuable information [14]? If an individual wants to use Amazon’s persuasion profile to jump-start a mobile exercise coach’s adaptation, there may or may not be technical and/or legal mechanisms to obtain and transfer this profile.
Non-disclosure
Means-adaptive persuasive systems are able and likely to not disclose that they are adapting to individuals. This can be contrasted with end adaptation, in which it is often advantageous for the agent to disclose the adaption and potentially easy to detect. For example, when Amazon recommends books for an individual it makes clear that these are personalized recommendations — thus benefiting from effects of apparent personalization and enabling presenting reasons why these books were recommended. In contrast, with means adaptation, not only may the results of the adaptation be less visible to users (e.g. emphasizing either “Pulitzer Prize winning” or “International bestseller”), but disclosure of the adaptation may reduce the target attitude or behavior change.
It is hypothesized that the effectiveness of social influence strategies is, at least partly, caused by automatic processes. According to dual-process models [4], un- der low elaboration message variables manipulated in the selection of influence strategies lead to compliance without much thought. These dual-process models distinguish between central (or systematic) processing, which is characterized by elaboration on and consideration of the merits of presented arguments, and pe- ripheral (or heuristic) processing, which is characterized by responses to cues as- sociated with, but peripheral to the central arguments of, the advocacy through the application of simple, cognitively “cheap”, but fallible rules [13]. Disclosure of means adaptation may increase elaboration on the implementations of the selected influence strategies, decreasing their effectiveness if they operate primarily via heuristic processing. More generally, disclosure of means adaptation is a disclosure of persuasive intent, which can increase elaboration and resistance to persuasion.
Implications of non-disclosure. The fact that persuasion profiles can be obtained and used without disclosing this to users is potentially a cause for concern. Potential reductions in effectiveness upon disclosure incentivize system designs to avoid disclosure of means adaptation.
Non-disclosure of means adaptation may have additional implications when combined with value being placed on the construction of an accurate persuasion profile. This requires some explanation. A simple system engaged in persuasion profiling could select influence strategies and implementations based on which is estimated to have the largest effect in the present case; the model would thus be engaged in passive learning. However, we anticipate that systems will take a more complex approach, employing active learning techniques [e.g., 6]. In active learning the actions selected by the system (e.g., the selection of the influence strategy and its implementation) are chosen not only based on the value of any resulting attitude or behavior change but including the value predicted improvements to the model resulting from observing the individual’s response. Increased precision, generality, or comprehensiveness of a persuasion profile may be valued (a) because the profile will be more effective in the present context or (b) because a more precise profile would be more effective in another context or more valuable in a market for persuasion profiles.
These later cases involve systems taking actions that are estimated to be non-optimal for their apparent goals. For example, a mobile exercise coach could present a message that is not estimated to be the most effective in increasing overall activity level in order to build a more precise, general, or comprehensive persuasion profile. Users of such a system might reasonably expect that it is designed to be effective in coaching them, but it is in fact also selecting actions for other reasons, e.g., selling precise, general, and comprehensive persuasion profiles is part of the company’s business plan. That is, if a system is designed to value constructing a persuasion profile, its behavior may differ substantially from its anticipated core behavior.
[1] Aarts, E.H.L., Markopoulos, P., Ruyter, B.E.R.: The persuasiveness of ambient intelligence. In: Petkovic, M., Jonker, W. (eds.) Security, Privacy and Trust in Modern Data Management. Springer, Heidelberg (2007)
[2] Baker, S.: Learning, and profiting, from online friendships. BusinessWeek 9(22) (May 2009)Selecting Effective Means to Any End 93
[3] Berdichevsky, D., Neunschwander, E.: Toward an ethics of persuasive technology. Commun. ACM 42(5), 51–58 (1999)
[4] Cacioppo, J.T., Petty, R.E., Kao, C.F., Rodriguez, R.: Central and peripheral routes to persuasion: An individual difference perspective. Journal of Personality and Social Psychology 51(5), 1032–1043 (1986)
[5] Cialdini, R.: Influence: Science and Practice. Allyn & Bacon, Boston (2001)
[6] Cohn,D.A., Ghahramani,Z.,Jordan,M.I.:Active learning with statistical models. Journal of Artificial Intelligence Research 4, 129–145 (1996)
[7] Eckles, D.: Redefining persuasion for a mobile world. In: Fogg, B.J., Eckles, D. (eds.) Mobile Persuasion: 20 Perspectives on the Future of Behavior Change. Stanford Captology Media, Stanford (2007)
[8] Fogg, B.J.: Persuasive Technology: Using Computers to Change What We Think and Do. Morgan Kaufmann, San Francisco (2002)
[9] Fogg, B.J.: Protecting consumers in the next tech-ade, U.S. Federal Trade Commission hearing (November 2006), http://www.ftc.gov/bcp/workshops/techade/pdfs/transcript_061107.pdf
[10] Fogg,B.J.:The behavior grid: 35 ways behavior can change. In: Proc. of Persuasive Technology 2009, p. 42. ACM, New York (2009)
[11] Kaptein, M., Aarts, E.H.L., Ruyter, B.E.R., Markopoulos, P.: Persuasion in am- bient intelligence. Journal of Ambient Intelligence and Humanized Computing 1, 43–56 (2009)
[12] Lacroix, J., Saini, P., Goris, A.: Understanding user cognitions to guide the tai- loring of persuasive technology-based physical activity interventions. In: Proc. of Persuasive Technology 2009, vol. 350, p. 9. ACM, New York (2009)
[13] Petty, R.E., Wegener, D.T.: The elaboration likelihood model: Current status and controversies. In: Chaiken, S., Trope, Y. (eds.) Dual-process theories in social psychology, pp. 41–72. Guilford Press, New York (1999)
[14] Prabhaker, P.R.: Who owns the online consumer? Journal of Consumer Market- ing 17, 158–171 (2000)
[15] Rawls, J.: The independence of moral theory. In: Proceedings and Addresses of the American Philosophical Association, vol. 48, pp. 5–22 (1974)
[16] Rhoads, K.: How many influence, persuasion, compliance tactics & strategies are there? (2007), http://www.workingpsychology.com/numbertactics.html
[17] Schafer, J.B., Konstan, J.A., Riedl, J.: E-commerce recommendation applications. Data Mining and Knowledge Discovery 5(1/2), 115–153 (2001)
- We were of course influenced by B.J. Fogg’s previous use of the term ‘persuasion profiling’, including in his comments to the Federal Trade Commission in 2006. [↩]
- This point can also be made in the language of interaction effects in analysis of variance: Persuasion profiles are estimates of person–strategy interaction effects. Thus, the end-independence of persuasion profiles requires not that the two-way strategy– context interaction effect is small, but that the three-way person–strategy–context interaction is small. [↩]
Adjusting biased samples
Nate Cohn at The New York Times reports on how one 19-year-old black man is having an outsized impact on the USC/LAT panel’s estimates of support for Clinton in the U.S. presidential election. It happens that the sample doesn’t have enough other people with similar demographics and voting history (covariates) to this panelist, so he is getting a large weight in computing the overall averages for the populations of interest, such as likely voters:
There is a 19-year-old black man in Illinois who has no idea of the role he is playing in this election.
He is sure he is going to vote for Donald J. Trump.
And he has been held up as proof by conservatives — including outlets like Breitbart News and The New York Post — that Mr. Trump is excelling among black voters. He has even played a modest role in shifting entire polling aggregates, like the Real Clear Politics average, toward Mr. Trump.
As usual, Andrew Gelman suggests that the solution to this problem is a technique he calls “Mr. P” (multilevel regression and post-stratification). I wanted to comment on some practical tradeoffs among common methods. Maybe these are useful notes, which can be read alongside another nice piece by Nate Cohn on how different adjustment methods can yield very different polling results.
Post-stratification
Complete post-stratification is when you compute the mean outcome (e.g., support for Clinton) for each stratum of people, such as 18-24-year-old black men, defined by the covariates X. Then you combine these weighting by the size of each group in the population of interest. This really only works when you have a lot of data compared with the number of strata — and the number of strata grows very fast in the number of covariates you want to adjust for.
Modeling sample inclusion and weighting
When people talk about survey weighting, often what they mean is weighting by inverse of the estimated probability of inclusion in the sample. You model selection into the survey S using, e.g., logistic regression on the covariates X and some interactions. This can be done with regularization (i.e., priors, shrinkage) since many of the terms in the model might be estimated with very few observations. Especially without enough regularization, this can result in very large weights when you don’t have enough of some particular type in your sample.
Modeling the outcome and integrating
You fit a model predicting the response (e.g., support for Clinton) Y with the covariates X. You regularize this model in some way so that the estimate for each person is going to “borrow strength” from other people with similar Xs. So now you have a fitted responses Yhat for each unique X. To get an estimate for a particular population of interest, integrate out over the distribution of X in that population. Gelman’s preferred version “Mr. P” uses a multilevel (aka hierarchical Bayes, random effects) model for the outcome, but other regularization methods may often be appealing.
This is nice because there can be some substantial efficiency gains (i.e. more precision) by making use of the outcome information. But there are also some practical issues. First, you need a model for each outcome in your analysis, rather than just having weights you could use for all outcomes and all recodings of outcomes. Second, the implicit weights that this process puts on each observation can vary from outcome to outcome — or even for different codings (i.e. a dichotomization of answers on a numeric scale) of the same outcome. In a reply to his post, Gelman notes that you would need a different model for each outcome, but that some joint model for all outcomes would be ideal. Of course, the latter joint modeling approach, while appealing in some ways (many statisticians love having one model that subsumes everything…) means that adding a new outcome to analysis would change all prior results.
Side note: Other methods, not described here, also work towards the aim of matching characteristics of the population distribution (e.g., iterative proportional fitting / raking). They strike me as overly specialized and not easy to adapt and extend.
A deluge of experiments
The Atlantic reports on the data deluge and its value for innovation.1 I particularly liked how Erik Brynjolfsson and Andrew McAfee, who wrote the Atlantic piece, highlight the value of experimentation for addressing causal questions — and that many of the questions we care about are causal.2
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.
- 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. [↩]
- 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. [↩]
- Note that two single-factor experiments over the same population with independent random assignment can be regarded as a single experiment with two factors. [↩]
Traits, adaptive systems & dimensionality reduction
Psychologists have posited numerous psychological traits and described causal roles they ought to play in determining human behavior. Most often, the canonical measure of a trait is a questionnaire. Investigators obtain this measure for some people and analyze how their scores predict some outcomes of interest. For example, many people have been interested in how psychological traits affect persuasion processes. Traits like need for cognition (NFC) have been posited and questionnaire items developed to measure them. Among other things, NFC affects how people respond to messages with arguments for varying quality.
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:
- obtain the questionnaire measure of the trait for a sample,
- predict this measure with data available for many individuals (e.g., log data),
- 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.
- I owe some clarity on this to some conversations with Mike Nowak and Maurits Kaptein. [↩]
Aardvark’s use of Wizard of Oz prototyping to design their social interfaces
The Wall Street Journal’s Venture Capital Dispatch reports on how Aardvark, the social question asking and answering service recently acquired by Google, used a Wizard of Oz prototype to learn about how their service concept would work without building all the tech before knowing if it was any good.
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.