=Paper=
{{Paper
|id=None
|storemode=property
|title=Building Persuasion Profiles in the Wild: Using Mobile Devices as Identifiers
|pdfUrl=https://ceur-ws.org/Vol-722/paper5.pdf
|volume=Vol-722
}}
==Building Persuasion Profiles in the Wild: Using Mobile Devices as Identifiers==
https://ceur-ws.org/Vol-722/paper5.pdf
Building Persuasion Profiles in the Wild: Using Mobile
Devices as Identifiers.
Maurits Kaptein
Eindhoven University of Technology / Philips Research
Den Dolech 2, 5600MB, the Netherlands. m.c.kaptein@tue.nl
ABSTRACT make their choice—the current version does not imple-
Tailoring — presenting the right message at the right ment social influence strategies and does not adapt to
time — has long been identified as one of the core op- its users.
portunities of persuasive systems. In this paper we de-
scribe a scenario in which an adaptive persuasive sys- Social Influence Strategies
tem which identifies users by the Bluetooth key of their Cialdini [2] shows how small changes to messages—such
mobile phone is used to promote energy savings. By as the message on the door—can increase their effective-
describing this simplistic system and its possible imple- ness. For example, a message in a hotel room asking
mentation we identify several key-criteria of adaptive guests to “reuse their towels” compared to a message
persuasive systems. stating “Join your fellow citizens in helping to save the
environment” led to a difference in towel re-usage of
Author Keywords
28.4% [7]. To structure thes types of messages Cialdini
[2] identifies six social influence strategies: Authority,
Persuasive Technology, Influence strategies
Consensus, Reciprocity, Liking, Scarcity, and Commit-
ment. The message in the towel re-usage example im-
ACM Classification Keywords plements the Consensus strategy: people act like other
H.1.2 User/Machine Systems: Software psychology. people do. A message (e.g.) stating that “The general
manager of this hotel requests you to re-use...” would
INTRODUCTION
implement the Authority strategy. These social influ-
ences strategies can easily be used to improve upon the
CHI2010 attendees were presented with a choice on
effectiveness of the paper-sign.
entering the conference hotel: A large revolving door
provided access to the hotel while next to it was a slid-
The final promise of persuasive technologies however—
ing door—some things simply do not fit through a re-
adapting influence attempts to individuals—will
volving door. With the air conditioning in full opera-
require some kind of interactive system. While
tion revolving doors are efficient at keeping the heat in.
adaptation of persuasive strategies to responses by
Sliding doors, however, are not. To help save energy a
users is mentioned early on in the literature on
paper-sign was put up: “Please take the revolving door”.
persuasive technologies Fogg [5, e.g.] we are unaware
A brief observation proved the paper-sign to be effec-
of any actual implementations.
tive just over half the time: 60% of the visitors took
the revolving door. This scenario, the “Revolving Door
Individual Differences
Problem”, offers a framework to describe adaptive per-
suasive systems. By further elaborating this scenario There is growing evidence that individuals differ in their
and exploring a solution we describe the neccesities and responses to influence strategies: Constructs like Need
difficulties that arise when designing adaptive persua- For Cognition [1] predict the response of individuals to
sive systems. the usage of social influence strategies. More concretely,
Kaptein et al. [9] show that usage of influence strategies
for individuals who are low susceptible to these strate-
The Promises of Persuasive Technology gies can lead to backfiring: for a portion of participants
There are three reasons why employing a persuasive sys- in their study compliance to a request was lower when
tem might be more effective than the current paper- the social influence strategy was presented. Next to this
sign: (1) Persuasive technologies function as social ac- overall tendency to respond to influence strategies, some
tors and can use social influence strategies, (2) they can individuals seem more likely to respond to one specific
be context aware, and (3) they can adapt to individual strategy—e.g. an authority argument—while others are
users [5, 8]. While the paper-sign is probably located more influenced by implementations of other strategies.
at the right place and at the right time—when visitors Cialdini et al. [3] shows that there are sizable and stable
individual differences in people’s responses to the com-
mitment strategy. Similar results have been obtained
when looking at the consensus strategy: Self-reported
susceptibility to this strategy highly correlates with be- of approaches the visitor has made to the doors and p
havioral responses to this strategy [10]. denotes the probability of success: the probability of
taking the revolving door. Given M messages one can
These individual differences in susceptibility to differ- compute for each individual, for each message, proba-
ent persuasive strategies imply that persuasive systems bility pm = km /nm where km is the number of observed
should personalize the way in which they attempt to successes after representation of message m, nm times
influence individuals. Such a class of systems, which we to a specific visitor. It makes intuitive sense to present
call adaptive persuasive systems, are an unexplored area a visitor with the messages with the highest pm .
in that we still need to understand how to model, design
and build these systems. This paper takes a concrete For a large number of observations N of one visitor this
but simple example that encapsulates the quintessence would make perfect sense. However, this will not inform
of this problem to discuss how to address these chal- a decision for a newly observed visitor. For a new visitor
lenges. one would present the message m for which pm is max-
imized over previously observed visitors1 . Actually—
SOLVING THE REVOLVING DOOR PROBLEM? given Stein’s result [4]—for every user a weighted aver-
Returning to the revolving door problem, let us consider age of the pm for an individual user and those of other
what is involved in implementing an adaptive persuasive users—one where the estimated p�m for an individual is
system. We need to (A) identify the visitors entering the “shrunk” toward the population mean—will provide a
lobby—minimally by giving each a unique ID, and (B) better estimate than an estimate based on observations
measure the effectiveness of a presented message. The of a single visitor alone. E.g., if the authority message is
Bluetooth key of visitor’s mobile phone could be used effective 70% of the time over all visitors and only 30%
for identification [11]. This will capture around 12% of percent of the time for the specific visitor under consid-
the visitors entering the lobby. This same identification eration, the best estimate of the (real) effectiveness of
method can also be used to measure the effectiveness of the authority message p�A for this visitor is a weighted
each persuasive attempt: One Bluetooth scanner next average of these two.
to the revolving door and one next to the sliding door
could determine which entrance was used by the current Adapting to Individuals
visitor. Based on this knowledge about the visitor and To include both the known effectiveness of a message
records of earlier decisions a message implementing the for others, and a specific visitors previous responses to
right influence strategy can be selected. In the remain- that same message, into a new estimate of message ef-
der of this paper, we focus on the mechanism by which fectiveness, pm , we use a Bayesian approach. A com-
these strategies can be selected. mon way of including prior information in a binomial
random process is to use the Beta-Binomial model [12].
Suppose we have only two messages to show, one imple- The Beta Beta(α, β) distribution functions as a con-
menting the authority strategy—“The general manager jugate prior to the binomial. If we re-parametrize the
of this hotel urges you to...” (A)—and one implement- beta distribution as follows
ing the consensus strategy—“80% of our visitors always
use...etc.” (B). The system then needs a mechanism to π(θ|µ, M ) = Beta(µ, M )
choose the message that is most likely to be effective
for the current visitor. It is intuitive that for a new vis- where µ = α+β α
and M = α + β, then the expected
itor the system should present the message which has value of the distribution is given by: E(θ|µ, M ) = µm .
lead to the highest compliance for other, previously ob- In our scenario this represents the expected probability
served, visitors. If this message is successful then there of a successful influence attempt by a specific message.
is no need to try different messages on subsequent visits. The certainty of this estimated success probability is
However, when the selected message is not effective, it represented by:
might become attractive to present another message on
the next visit. This decision logically depends on the µ(1 − µ)
V ar(θ|µ, M ) = σ 2 =
initial succes probabilities of the messages under con- M +1
sideration, the variance of effectiveness of messages be-
tween visitors, and the number of succes’s or failures ob- After specifying the probability of success µm of mes-
served for the current visitor. A collection of estimates sage m and the certainty about this estimate σm 2
we can
of the effectiveness of different influence strategies for treat this as our prior expectancy about the effective-
an individual is called a Persuasion Profile and can be ness of a specific message and update this expectancy
used to select the most-likely-to-be effective message on by multiplying it by the likelihood of the observation(s)
a next visit. to obtain the distribution of our posterior expectation:
p(θ|k) ∝ l(k|θ)π(θ|µ, M )
Formalizing the Adaptation Problem = Beta(k + M µ, n − k + M (1 − µ))
The probability of a single visitor taking the revolving
door on multiple occasions can be regarded a binomial 1
This is assuming the error costs—the effects of presenting
random variable B(n, p) where n denotes the number the wrong message—are equal for each message.
The newly obtained Beta distribution, B(µ, M ), func- A and 50% to strategy B, (2) susceptible visitors,
tions as our probability distribution with a new point- A = 40%, B = 90%, (3) visitors susceptible to message
estimate of the effectiveness of the presented message B, A = 10%, B = 90%, and (4) visitors susceptible
given by: to message A, A = 90%, B = 10%. Table 1 shows
k + Mµ an excerpt of the simulated data. Based on these
E(θ|k) = simulated data we first compute our population
n+M estimates of message effectiveness for each message:
Decision Rule p�A = 0.38, p�B = 0.58. Thus, message B—the
The Beta-Binomial model described above allows us to consensus message—was most effective.
estimate the effectiveness of message m, include prior
knowledge, and update these estimates based on new Type User Occasion Mes. A Mes. B
observations. A individual’s persuasion profile would 1 1 1 1 0 0
be a record of both the expected success, µm , and the 2 1 1 2 0 0
certainty, σm
2
of different influence strategies. 3 1 1 3 0 1
.. .. .. .. .. ..
To determine which message to present next, one could .. .. .. .. .. ..
pick the message which has the highest µm . However, if 1000 4 20 50 1 0
σm2
is large this decision might not be feasible given that
Table 1. Overview of the simulated data for the 4 differ-
the difference between effectiveness estimates might not ent user groups. Columns Mes. A and Mes. B represent
be significant. To address this we can choose to show the success of the influence message at that point in time.
the message with the highest estimate when this es-
timate is “certain enough”—in the binomial case only Next, we simulate for each visitor, each occurrence at
once sufficient observations are obtained. In uncertain the doors. We select the message as specified by our
situations we can randomly present one of the H mes- decision rule and record the (simulated) outcome. Next,
sages which have the highest estimates out of the total we update our expectancy for the selected message and
set of estimates of M messages. This decision rule would iterate through all occurrences. To ensure a flexible
avoid presenting each new visitor with only the single starting point for each user we set the prior variance of
most effective message when responses to messages are each estimate at the first encounter to be high: σA 2
=
variant. σB = 0.05.
2 3
Because the Beta distribution is not necessarily sym- Figure 1 shows for four users—one out of each group—
metrical the variance σm
2
provides and inadequate start- in separate panels, the estimated probability of success
ing point to compute confidence intervals. This prob- of the two messages (left and right side of each panel).
lem can be solved using simulations: By generating a In the upper left panel—representing a general insus-
number of draws from the specified Beta distribution ceptible visitor —convergence to message B, whose esti-
and computing (e.g.) the 20th and 80th percentiles one mated effect is presented on the right side of the upper
can compute a empirical confidence interval. The above left panel, is slow: it takes about 40 observations before
described decision rule for M = 2 would then result in: B is consistently estimated to be the “best” message.
1 µ1 > P erc(80)2 With higher compliance and/or larger differences in ef-
Mselected = 2 µ2 > P erc(80)1 fectiveness of the two strategies convergence is much
faster. The bottom right of figure 1 shows a user from
Rand(1, 2) otherwise the visitors susceptible to message A group. For this
Thus, if the estimated effectiveness of a message 1, user after 10 observations strategy A is correctly iden-
p�1 = µ1 , is higher than the 80th percentile of message tified as the most successful strategy.
2, P erc(80)2 , the system presents message one.2 If the
confidence interval of two messages overlap the system Limitations of the proposed solution
could randomly present one of these two. There are a number of drawbacks of the proposed
Beta-Binomial solution to create adaptive persuasive
Simulations systems. Besides the fact that when the number of
To explore the presented Beta-Binomial approach in strategies grows the number of necessary occasions
the M = 2 scenario we simulated a dataset presenting for convergence will increase, there are three more
different visitors observed at multiple points in time. fundamental issues which are not addressed by this
The simulated data describes the message success algorithm. First, while including prior information
of two different messages for four different groups based on other users, the algorithm described here
of visitors with 20 visitors each on 50 approaches does not use a shrunken estimate on each occasion:
to the doors. The four groups represent (1) general After including the initial knowledge of the behavior
insusceptible visitors—those that respond favorable to of other visitors the model is specific for an individual
only 10% of the message which implement strategy 3
One could estimate this variance based on the between-
2
The 80th percentile is an arbitrary choice. visitor variance.
0 1020304050 0 1020304050 Mobile devices—as used in our scenario—provide a core
opportunity to serve as an identifier for adaptive persua-
lower + mean + upper
lower + mean + upper
0.8 0.8 sive technologies. Currently we are operating a system,
0.6 0.6 like the one described here, in real-life and we would
0.4 0.4 like to share our experiences building and deploying this
system during the CHI 2011 PINC workshop.
0.2 0.2
0 1020304050 0 1020304050
References
Occasions Occasions
[1] Cacioppo, J. T. and Petty, R. E. (1982). The need
for cognition. J. of Pers. and Soc. Psy.
0 1020304050 0 1020304050 [2] Cialdini, R. (2001). Influence, Science and Practice.
Allyn & Bacon, Boston.
lower + mean + upper
lower + mean + upper
0.8 0.8
[3] Cialdini, R. B., Trost, M. R., and Newsom, J. T.
0.6 0.6 (1995). Preference for consistency: The development
0.4 0.4 of a valid measure and the discovery of surprising
0.2 0.2
behavioral implications. J. of Pers. and Soc. Psy.,
69:318–328.
0 1020304050 0 1020304050
[4] Efron, B. and Morris, C. (1975). Data analysis us-
Occasions Occasions ing Stein’s estimator and its generalizations. Journal
of the American Statistical Association, 70(350):311–
Figure 1. Progression of point estimates of the effects of 319.
two messages on four different users (the four panels).
Within each panel the left side shows the estimated ef- [5] Fogg, B. J. (2002). Persuasive Technology: Using
fect of message A, including in gray its 80% confidence
interval, and the right shows the estimates for message
Computers to Change What We Think and Do. Mor-
B. A horizontal section in the estimates of message A gan Kaufmann.
indicates that at that point in time the message B was
shown and updated. [6] Gelman, A. and Hill, J. (2007). Data Analysis Using
Regression and Multilevel/Hierarchical Models. Cam-
bridge University Press.
visitor. While this provides quick adaptation there
is no opportunity to adapt estimates based on [7] Goldstein, N. J., Cialdini, R. B., and Griskevicius,
changing population wise trends. Second, since V. (2008). A room with a viewpoint: Using social
the estimates for the effectiveness of the strategies norms to motivate environmental conservation in ho-
are treated independently there is no way to of tels. J. of Cons. Res., 35(3):472–482.
“borrowing strength” [6] based on correlations with [8] Kaptein, M., Aarts, E. H. L., Ruyter, B. E. R., and
other strategies. Both of these concerns could be Markopoulos, P. (2009a). Persuasion in ambient in-
addressed using a multilevel approach. Finally, the telligence. Journal of Ambient Intelligence and Hu-
proposed model provides no method of including prior manized Computing, 1:43—56.
believes about the distribution of visitor profiles over a
population. [9] Kaptein, M., Lacroix, J., and Saini, P. (2010). Indi-
vidual differences in persuadability in the health pro-
CONCLUSIONS motion domain. In Ploug, T., Hasle, P., and Oinas-
We identified two core necessities of adaptive persua- Kukkonen, H., editors, Persuasive Technology, pages
sive systems: a means to identify users and a means to 94—105. Springer Berlin / Heidelberg.
measure effectiveness of persuasive attempts. Further- [10] Kaptein, M. C., de Ruyter, B., Markopoulos, P.,
more, we highlighted a number of challenges associated and Aarts, E. (2009b). Persuading you: Individual
with the design of these systems. The presented Beta- differences in susceptibility to persuasion. In 12th
Binomial solution is lightweight and functions well in IFIP TC 13 International Conference - INTERACT,
simulations with only two messages. More elaborate pages 24–28, Uppsala, Sweden. ACM Press.
algorithms which are (1) variant to changing popula-
tion trends,(2) allow for relationships between strate- [11] Kostakos, V. (2008). Using bluetooth to capture
gies, and (3) enable us to include prior beliefs about passenger trips on public transport buses. CoRR,
user profiles should be explored. Given the current state abs/0806.0874.
of social science literature on influence strategies we be-
lieve that persuasive technologies should tailor the influ- [12] Wilcox, R. R. (1981). A review of the beta-binomial
ence strategies they use to their users. We described one model and its extensions. J. of Educ. Stat., 6:3–32.
possible—but limited—implementation of such a sys-
tem. This, and other, implementations should now be
tested empirically.