Tailoring - presenting the right message at the right time - has long been identified as one of the core opportunities of persuasive systems. In this paper we describe a scenario in which an adaptive persuasive system which identifies users by the Bluetooth key of their mobile phone is used to promote energy savings. By describing this simplistic system and its possible implementation we identify several key-criteria of adaptive persuasive systems.
INTRODUCTION CHI2010 attendees were presented with a choice on entering the conference hotel: A large revolving door provided access to the hotel while next to it was a sliding door—some things simply do not fit through a revolving door. With the air conditioning in full operation revolving doors are efficient at keeping the heat in. Sliding doors, however, are not. To help save energy a paper-sign was put up: “Please take the revolving door”. A brief observation proved the paper-sign to be effective just over half the time: 60% of the visitors took the revolving door. This scenario, the “Revolving Door Problem”, offers a framework to describe adaptive persuasive systems. By further elaborating this scenario and exploring a solution we describe the neccesities and difficulties that arise when designing adaptive persuasive systems.
The Promises of Persuasive Technology
There are three reasons why employing a persuasive
system might be more effective than the current
papersign: (1) Persuasive technologies function as social
actors and can use social influence strategies, (2) they can
be context aware, and (3) they can adapt to individual
users [
Social Influence Strategies
Cialdini [
The final promise of persuasive technologies however— adapting influence attempts to individuals—will require some kind of interactive system. While adaptation of persuasive strategies to responses by users is mentioned early on in the literature on persuasive technologies Fogg [5, e.g.] we are unaware of any actual implementations.
Individual Differences
There is growing evidence that individuals differ in their
responses to influence strategies: Constructs like Need
For Cognition [
These individual differences in susceptibility to different persuasive strategies imply that persuasive systems should personalize the way in which they attempt to influence individuals. Such a class of systems, which we call adaptive persuasive systems, are an unexplored area in that we still need to understand how to model, design and build these systems. This paper takes a concrete but simple example that encapsulates the quintessence of this problem to discuss how to address these challenges.
SOLVING THE REVOLVING DOOR PROBLEM?
Returning to the revolving door problem, let us consider
what is involved in implementing an adaptive persuasive
system. We need to (A) identify the visitors entering the
lobby—minimally by giving each a unique ID, and (B)
measure the effectiveness of a presented message. The
Bluetooth key of visitor’s mobile phone could be used
for identification [
Suppose we have only two messages to show, one implementing the authority strategy—“The general manager of this hotel urges you to...” (A)—and one implementing the consensus strategy—“80% of our visitors always use...etc.” (B). The system then needs a mechanism to choose the message that is most likely to be effective for the current visitor. It is intuitive that for a new visitor the system should present the message which has lead to the highest compliance for other, previously observed, visitors. If this message is successful then there is no need to try different messages on subsequent visits. However, when the selected message is not effective, it might become attractive to present another message on the next visit. This decision logically depends on the initial succes probabilities of the messages under consideration, the variance of effectiveness of messages between visitors, and the number of succes’s or failures observed for the current visitor. A collection of estimates of the effectiveness of different influence strategies for an individual is called a Persuasion Profile and can be used to select the most-likely-to-be effective message on a next visit.
Formalizing the Adaptation Problem
The probability of a single visitor taking the revolving
door on multiple occasions can be regarded a binomial
random variable B(n, p) where n denotes the number
of approaches the visitor has made to the doors and p
denotes the probability of success: the probability of
taking the revolving door. Given M messages one can
compute for each individual, for each message,
probability pm = km/nm where km is the number of observed
successes after representation of message m, nm times
to a specific visitor. It makes intuitive sense to present
a visitor with the messages with the highest pm.
For a large number of observations N of one visitor this
would make perfect sense. However, this will not inform
a decision for a newly observed visitor. For a new visitor
one would present the message m for which pm is
maximized over previously observed visitors1. Actually—
given Stein’s result [
Adapting to Individuals
To include both the known effectiveness of a message
for others, and a specific visitors previous responses to
that same message, into a new estimate of message
effectiveness, pm, we use a Bayesian approach. A
common way of including prior information in a binomial
random process is to use the Beta-Binomial model [
π(θ|µ, M ) = Beta(µ, M ) where µ = α+αβ and M = α + β, then the expected value of the distribution is given by: E(θ|µ, M ) = µm. In our scenario this represents the expected probability of a successful influence attempt by a specific message. The certainty of this estimated success probability is represented by:
V ar(θ|µ, M ) = σ2 = µ(1 − µ) M + 1 After specifying the probability of success µm of message m and the certainty about this estimate σm2 we can treat this as our prior expectancy about the effectiveness of a specific message and update this expectancy by multiplying it by the likelihood of the observation(s) to obtain the distribution of our posterior expectation: p(θ|k) ∝ l(k|θ)π(θ|µ, M ) =
Beta(k + M µ, n − k + M (1 − µ)) 1This is assuming the error costs—the effects of presenting the wrong message—are equal for each message. The newly obtained Beta distribution, B(µ, M ), functions as our probability distribution with a new pointestimate of the effectiveness of the presented message given by:
E(θ|k) = k + M µ n + M Decision Rule The Beta-Binomial model described above allows us to estimate the effectiveness of message m, include prior knowledge, and update these estimates based on new observations. A individual’s persuasion profile would be a record of both the expected success, µm, and the certainty, σm2 of different influence strategies. To determine which message to present next, one could pick the message which has the highest µm. However, if σm2 is large this decision might not be feasible given that the difference between effectiveness estimates might not be significant. To address this we can choose to show the message with the highest estimate when this estimate is “certain enough”—in the binomial case only once sufficient observations are obtained. In uncertain situations we can randomly present one of the H messages which have the highest estimates out of the total set of estimates of M messages. This decision rule would avoid presenting each new visitor with only the single most effective message when responses to messages are variant.
Because the Beta distribution is not necessarily symmetrical the variance σm2 provides and inadequate starting point to compute confidence intervals. This problem can be solved using simulations: By generating a number of draws from the specified Beta distribution and computing (e.g.) the 20th and 80th percentiles one can compute a empirical confidence interval. The above described decision rule for M = 2 would then result in: Mselected = 1 2 µ1 > P erc(80)2 µ2 > P erc(80)1 Rand(1, 2) otherwise Thus, if the estimated effectiveness of a message 1, p1 = µ1, is higher than the 80th percentile of message 2, P erc(80)2, the system presents message one.2 If the confidence interval of two messages overlap the system could randomly present one of these two.
Simulations To explore the presented Beta-Binomial approach in the M = 2 scenario we simulated a dataset presenting different visitors observed at multiple points in time. The simulated data describes the message success of two different messages for four different groups of visitors with 20 visitors each on 50 approaches to the doors. The four groups represent (1) general insusceptible visitors —those that respond favorable to only 10% of the message which implement strategy 2The 80th percentile is an arbitrary choice.
A and 50% to strategy B, (2) susceptible visitors, A = 40%, B = 90%, (3) visitors susceptible to message B, A = 10%, B = 90%, and (4) visitors susceptible to message A, A = 90%, B = 10%. Table 1 shows an excerpt of the simulated data. Based on these simulated data we first compute our population estimates of message effectiveness for each message: pA = 0.38, pB = 0.58. Thus, message B—the consensus message—was most effective. Next, we simulate for each visitor, each occurrence at the doors. We select the message as specified by our decision rule and record the (simulated) outcome. Next, we update our expectancy for the selected message and iterate through all occurrences. To ensure a flexible starting point for each user we set the prior variance of each estimate at the first encounter to be high: σA2 = σB2 = 0.05.3 Figure 1 shows for four users—one out of each group— in separate panels, the estimated probability of success of the two messages (left and right side of each panel). In the upper left panel—representing a general insusceptible visitor —convergence to message B, whose estimated effect is presented on the right side of the upper left panel, is slow: it takes about 40 observations before B is consistently estimated to be the “best” message. With higher compliance and/or larger differences in effectiveness of the two strategies convergence is much faster. The bottom right of figure 1 shows a user from the visitors susceptible to message A group. For this user after 10 observations strategy A is correctly identified as the most successful strategy.
Limitations of the proposed solution There are a number of drawbacks of the proposed Beta-Binomial solution to create adaptive persuasive systems. Besides the fact that when the number of strategies grows the number of necessary occasions for convergence will increase, there are three more fundamental issues which are not addressed by this algorithm. First, while including prior information based on other users, the algorithm described here does not use a shrunken estimate on each occasion: After including the initial knowledge of the behavior of other visitors the model is specific for an individual 3One could estimate this variance based on the betweenvisitor variance. 0 1020304050
Occasions 0 1020304050
Occasions
visitor. While this provides quick adaptation there
is no opportunity to adapt estimates based on
changing population wise trends. Second, since
the estimates for the effectiveness of the strategies
are treated independently there is no way to of
“borrowing strength” [
CONCLUSIONS We identified two core necessities of adaptive persuasive systems: a means to identify users and a means to measure effectiveness of persuasive attempts. Furthermore, we highlighted a number of challenges associated with the design of these systems. The presented BetaBinomial solution is lightweight and functions well in simulations with only two messages. More elaborate algorithms which are (1) variant to changing population trends,(2) allow for relationships between strategies, and (3) enable us to include prior beliefs about user profiles should be explored. Given the current state of social science literature on influence strategies we believe that persuasive technologies should tailor the influence strategies they use to their users. We described one possible—but limited—implementation of such a system. This, and other, implementations should now be tested empirically.
0 1020304050
Occasions 0 1020304050
Occasions
Mobile devices—as used in our scenario—provide a core opportunity to serve as an identifier for adaptive persuasive technologies. Currently we are operating a system, like the one described here, in real-life and we would like to share our experiences building and deploying this system during the CHI 2011 PINC workshop.