=Paper=
{{Paper
|id=None
|storemode=property
|title=Decoy Effects in Financial Service E-Sales Systems
|pdfUrl=https://ceur-ws.org/Vol-811/paper1.pdf
|volume=Vol-811
}}
==Decoy Effects in Financial Service E-Sales Systems==
https://ceur-ws.org/Vol-811/paper1.pdf
Decoy Effects in Financial Service E-Sales Systems
Erich Christian Teppan Alexander Felfernig, Klaus Isak
Dept. of Applied Informatics Dept. of Software Technology
Alpen-Adria Universitaet Klagenfurt Graz Instititute of Technology
9020 Klagenfurt, Austria 8010 Graz, Austria
Erich.Teppan@uni.klu.ac.at Alexander.Felfernig@ist.tu-graz.at
Klaus.Isak@ist.tu-graz.at
ABSTRACT 1. INTRODUCTION
Users of E-Sales platforms typically face the problem of It is often hard for customers of E-sales platforms to find
choosing the most suitable product or service from large suitable products or services (denoted as items for the re-
and potentially complex assortments. Whereas the problem mainder of this paper) which match their requirements. This
of finding and presenting suitable items fulfilling the user’s challenge is triggered by the size and complexity of the un-
requirements can be tackled by providing additional support derlying item assortment. Recommender applications facil-
in the form of recommender- and configuration systems, the itate the item identification process by proactively support-
control of psychological side e↵ects resulting from irrational- ing the customer/user in di↵erent types of decision scenarios
ities of human decision making has been widely ignored so [7]. These systems have been a very active research field for
far. Decoy e↵ects are one family of biases which have been many years which resulted in di↵erent solutions for many
shown to be relevant in this context. The asymmetric dom- item domains [10][13][15][19]. What has been widely ignored
inance e↵ect and the compromise e↵ect have been shown to by the E-sales and recommender community is that once sets
be among the most stable decoy e↵ects and therefore also of items are presented on some sort of result page, decision
carry big potential for biasing online decision taking. This phenomena occur which can have significant impacts on cus-
paper presents two user studies investigating the impacts tomer decision making [30].
of the asymmetric dominance and compromise e↵ect in the One family of e↵ects which have been shown to be rele-
financial services domain. While the first study uses synthe- vant in this context are decoy e↵ects [30]. The decoy e↵ect
sized items for triggering a decoy e↵ect, the second study induces an increased attraction of target items with respect
uses real products found on konsument.at, which is an Aus- to competitor items due to the existence of so-called decoy
trian consumer advisory site. Whereas the results of the first items. In other words, the target items are those which
study prove the potential influence of decoy e↵ects on online (should) profit more from the existence of the decoys than
decision making in the financial services domain, the results the competitors. Two prominent types of decoy e↵ects are
of the second study provide clear evidence of the practical the asymmetric dominance e↵ect (ADE) [11] and the com-
relevance for real online decision support- and E-sales sys- promise e↵ect (CE) [31]. These two decoy e↵ects di↵er in
tems. terms of the relative positions (described by the correspond-
ing attribute dimensions – in our example: optical zoom and
resolution) of the decoy items in the item landscape (see
Categories and Subject Descriptors Figure 1). Compared to the target item, an asymmetrically
H.5.2 [INFORMATION INTERFACES AND PRE- dominated decoy item (see {d1, d2, d3} in Figure 1) is worse
SENTATION]: User Interfaces—Graphical user interfaces in every dimension (d1) or worse in at least one dimension
(GUI) and equal in the other dimensions (d2 and d3). Compared
to the competitor, the asymmetrically dominated decoy item
General Terms is - though worse in some dimensions - also better in some
dimensions. In other words, there are dimensions where the
Human Factors, Experimentation, Theory decoy item defeats the competitor, but the decoy defeats the
target in none of the given dimensions.
Keywords Table 1 shows a simplified example of the ADE (d1) with
Decision Phenomena, Decoy E↵ects, Consumer Decision Mak- two attribute dimensions and two items in the domain of
ing, E-Sales Platforms. digital cameras: The target item is better than the com-
petitor in the dimension resolution (8 mpix) whereas the
competitor is better in the dimension optical zoom (6x). In
theory, the addition of an asymmetrically dominated decoy
the attractiveness of the target increases.
The asymmetry induced by the decoy is most easily shown
by the corresponding domination graph which outlines the
superiority/inferiority relations between all items in every
dimension. Figure 2 is showing the corresponding domina-
Decisions@RecSys 2011, ChicagoLJ8 tion graphs for the example in Table 1.
1
Competitor Target Decoy
Resolution 10 mpix 7 mpix 3 mpix
Optical Zoom 3x 6x 7x
Table 2: Example of the CE in the domain of digital
camera. The additional presentation of the decoy
makes the target a good compromise.
In a set without decoy (a), the target and the competi-
tor items are dominating each other in the same number of
attributes (i.e. in our case in on attribute dimension each).
Due to the inclusion of a decoy item the situation changes
(b). Now the target dominates the rest of the set more than
the competitor (i.e. three arrows vs. two arrows). As a
direct consequence of asymmetrical dominance, d1, d2, and
d3 are inferior items such that the overall utility calculated
with some objective utility function (e.g. multi attribute
utility theory [32]) is lower compared to the target.
Figure 1: Asymmetrically dominated decoy items Another important decoy e↵ect is the compromise e↵ect
(area ADE) and decoy items triggering a compro- [24][31] (see d4 and d5 in Figure 1). The key reason for the
mise e↵ect (area CE) for the benefit of the target. existence of this e↵ect is the fact that consumers rather pre-
fer items with medium values in all dimensions than items
with extreme values (”good” compromise items). This aspect
of human choice behavior is denoted extremeness aversion
[28]. Table 2 shows a very simple example.
Again, by the addition of the compromise decoy (d4) the
attractivity of the target item is increased compared to the
Competitor Target Decoy attractivity of the competitor item. The distinction between
Resolution 8 mpix 5 mpix 6 mpix d4 and d5 is based on an objective utility function [23]. Hav-
Optical Zoom 3x 6x 2x ing such a utility function, all items positioned on the diag-
onal in Figure 1 are pareto optimal. As a consequence, a
Table 1: Example of the ADE in the domain of dig- d5-decoy has the same overall utility as the target, i.e. does
ital cameras. The additional presentation of the de- not constitute an inferior item. In this case the only mecha-
coy shifts the attraction towards the target. nism causing the compromise e↵ect is extremeness aversion.
As a d4-decoy also constitutes an extreme item, it triggers
extremeness aversion. Additionally it constitutes an inferior
item such that the occurring tradeo↵ contrasts support pos-
itive influences for the target. Tradeo↵ contrasts exist when
the advantages of one item outweigh the advantages of an-
other item. In the example of Table 2, the target is much
better in the dimension resolution than it is defeated by the
decoy in the dimension of optical zoom. As discussed above,
the extreme case of a tradeo↵ contrast leads to dominance.
The major contributions of this paper are the following:
We provide an in-depth analysis of the existence of decoy
e↵ects in the financial services domain. In this context
we show the existence of decoy e↵ects for result sets with
more than three items and also show the e↵ects on the ba-
sis of commercial product assortments. The investigations
concentrate on the two most important e↵ects, namely the
asymmetric dominance e↵ect and the compromise e↵ect. All
Figure 2: Domination graph: Without decoy (a) the presented studies have been carried out online and unsuper-
target as well as the competitor dominate each other vised and thus preserved a maximum of real world condi-
in one dimension. After inclusion of the decoy (b) tions. The results of the presented empirical studies clearly
the equality seems to change as the target dominates show the impact of decoy e↵ects on item selection behavior
the decoy in both dimensions whereas the competi- of users. Consequently, although not taken into account up
tor dominates the decoy in only one dimension and to now, these e↵ects play a major role for the construction
is even dominated by the decoy in one dimension of recommender and esales applications.
(blue/spotted arrow). The remainder of this paper is organized as follows. In
Section 2 we provide an overview of related work. In Sec-
tion 3 we discuss the results of a user study based on a
synthesized set of financial services. In the following (Sec-
2
tion 4) we present the results of the second decoy study tigating decoy e↵ects in realistic settings, as all studies are
which is based on a real-world dataset (bankbooks from kon- carried out unsupervised using a recommender like online
sument.at). The impact of decoy e↵ects on the construction system. The second study presented in this paper uses real
of recommender applications is summarized in Section 5. market data (i.e. real capital savings books) taken from an
With Section 6 we conclude the paper and provide an out- independent consumer information site (www.konsument.at).
look of future work. Moreover, financial services constitutes a high-involvement
decision domain, such that decoy e↵ects should be less likely
2. RELATED WORK than in low-involvement domains where the user does not
put too much energy into the decision process.
The main reason why decoy e↵ects occur in human deci-
sion making is that humans often do not act fully rational.
Fully rational agents apply some sort of value maximiza- 3. EXPERIMENT WITH SYNTHESIZED
tion model like the multi attribute utility theory, multiple SETS OF FINANCIAL SERVICES
regression, or Bayesian statistics in order to find an optimal
In order to investigate the influence of the Asymmetric
solution [23][25][32]. All these approaches are computation-
Dominance- and Compromise E↵ects (ADE and CE) on
ally very expensive, but human decision taking is normally
product selection tasks in the financial services domain, a
bounded by time restrictions, limited cognitive capacities,
corresponding online user study was carried out. The ex-
and limited willingness to accept cognitive e↵ort. This is the
periment was two folded: Subjects (Students of the Alpen-
reason why humans apply in many circumstances heuristic
Adria Universitaet Klagenfurt) had to accomplish one deci-
approaches (i.e. rules of thumb).
sion task for each e↵ect (Asymmetric Dominance- and Com-
In contrast to rationality, this concept is called bounded
promise E↵ect). Altogether there were 535 valid sessions
rationality or procedural rationality [12][22][26][27]. Gerd
whereby 358 were from female persons. The subject’s age
Gigarenzer has shown in multiple experiments, that heuris-
ranged from 18 to 76 years (mean = 25.8, std = 7.2).
tic, bounded rational approaches can be as accurate as some
fully rational concept like multiple regression [8][9]. Unfor- 3.1 Compromise Effect
tunately, there are cases where bounded rationality acts as a
door opener for systematic misjudgements which builds the Design
grounding for decision phenomena/e↵ects. Based on mis- The first decision task the subjects had to accomplish was
judgements due to bounded rationality, these decision e↵ects to decide which type of financial service they would choose
bear the danger of suboptimal decision making [30]. if they had 5000 Euros. Depending on the products the sub-
Decoy e↵ects [1][11][17][20][21][31] are one family of such jects had to choose from, three groups were di↵erentiated:
e↵ects which have the potential of severely impacting on the The control group with the product types public bonds, gold,
perceived value of goods and services. Basically, there exist mixed funds, group Decoy A with the product types bank-
three types of decoy e↵ects: the attraction e↵ect [21], the book (=decoy), public bonds, gold, mixed funds, and group
asymmetric dominance e↵ect [11], and the compromise e↵ect Decoy B containing the product types public bonds, gold,
[24][31]. In existing literature, the expressions decoy e↵ect, mixed funds, shares (=decoy) (see Figure 3).
asymmetric dominance e↵ect and attraction e↵ect are often
used synonymously, as the asymmetric dominance e↵ect is
the most prominent and stable decoy e↵ect, and the attrac-
tion e↵ect could be seen as the more general e↵ect sharing
the principle of tradeo↵ contrasts [28]. A clear distinction
between the di↵erent e↵ects, the corresponding decoy items,
and the di↵erent mechanisms working behind the di↵erent
decoy e↵ects can be found in [29]. Since the 1980’s a lot of
research has been done in order to investigate decoy e↵ects.
While the existence of such biases has been shown in quite
a number of publications there has not been done much re-
search in investigating the impacts of such decision biases in
real world sales platforms with realistic environments and
on the basis of real market data. This is out of two rea-
sons: First, the investigation of some decision e↵ect under
clean room conditions makes it possible to eliminate a max-
imum of disturbing influences and therefore also maximizes
and purifies the measured e↵ect. Second, it is not easy
to get good market data as companies are usually very re- Figure 3: Product landscape of CE-task: three al-
served concerning the proliferation of business intelligence. ternative items (funds, gold, bonds) with the cor-
Although the investigation of cognitive biases without real responding decoy items (shares and bankbook, re-
market conditions are indeed relevant from the basic re- spectively).
search point of view, the practical relevance for real world
applications cannot be assessed because a particular bias can The utility of each product was described in terms of risk
be too small in relation to other overlaying (uncontrolled) and return rate (see Figure 3 and Table 3), whereby low
e↵ects such that the practical relevance for real world appli- risk and a high return rate was interpreted as good (i.e.
cations is possibly not given. high utility value).
Closing this gap, this paper is in the line of research inves- As exact preference models were not given equal weighted
3
Product type Return Rate Risk were di↵erentiated: The control group contained the prod-
Bankbook 1 8
ucts bankbook1, bankbook2, bankbook3. Group Decoy 1 con-
Public bonds 4 6
Gold 5 5 tained bankbook1, bankbook2, bankbook3, decoy1. Group De-
Mixed funds 6 4 coy 2 contained bankbook1, bankbook2, bankbook3, decoy2,
Shares 8 1 and group Decoy 3 contained bankbook1, bankbook2, bank-
book3, decoy3 (see Figure 5). decoyX denotes an asymmet-
Table 3: Product utilities in CE-task. rically dominated decoy for bankbookX. When presenting the
items to the user, the decoy items (decoy1, decoy2, decoy3)
were called bankbook4 (in order to avoid experimental side
Multi attribute utility Theory (MAUT [32]) was used for de- e↵ects triggered by the item name).
signing suitable options. Although exact knowledge about
user preferences (e.g. attribute weights) would be preferred
also a linear equal weight model does the job as all hypothe-
ses are tested on behalf of corresponding control groups re-
vealing the actual preferences. The product types bonds,
gold, and funds have the same overall utility (= 10) and
therefore no tradeo↵ contrasts (TC) occur (see Table 3).
The extreme options bankbook and shares have a little lower
overall utility (= 9). Adding such options leads to TCs and
therefore can cause compromise e↵ects.
There were two hypotheses postulated:
• H1: Choice of Bonds is increased by the presence of
Bankbook.
• H2: Choice of Funds is increased by the presence of
Shares.
Results
Generally, users preferred low risk items over high return
items. Comparing the choice distribution of the control Figure 5: Product landscape of ADE-task: three
group with group Decoy A [H1], it can be said that more alternative bankbooks with the corresponding decoy
people chose bonds in the decoy group than in the control items.
group (see Figure 4). In fact, the presence of bankbook made
bonds the strongest option whereas in the control group gold The utility of each product was described in terms of inter-
was the most often chosen product type. The corresponding est rate per year (p.a.) and binding in months (i.e. the time
statistical analysis of bonds choices in the two groups showed within it is not possible to withdraw the money), whereby
a strong tendency (Fisher’s Exact Test, one-sided: p <.079). low binding and high interest rate was interpreted as good
Comparing the choice distribution of the control group and (i.e. high utility value). Figure 5 and Table 4 summarize
group Decoy B [H2] the e↵ect is even clearer. The increase the settings.
of funds choices in presence of shares was highly significant
(Fisher’s Exact Test, one-sided: p <.001). It is notable that Product Interest rate p.a. Binding in months
in all three groups the compromise options (i.e. the prod- Bankbook1 4.8 12
uct groups in the middle) scored better than the extreme Bankbook2 4.4 6
options. Bankbook3 4.0 0
Decoy1 4.7 12
Decoy2 4.3 6
Decoy3 3.9 0
Table 4: Product attributes in ADE-task.
In the control group no tradeo↵ contrasts (TCs) where ex-
istent as the product with the highest interest rate had also
the longest binding and vice versa. The decoy products (de-
coy1, decoy2, decoy3) constitute asymmetrically dominated
alternatives (e.g. decoy1 is only dominated by bankbook1,
Figure 4: Choice distribution for the CE-task. etc). Additionally to the ADE-constellation there can also
be found further TCs between the decoy and the non dom-
inating bankbooks (i.e. compromise e↵ects).
3.2 Asymmetric Dominance Effect There were three hypotheses postulated:
Design • H3: Choice of Bankbook1 is increased by the presence
In the second decision task the subjects also had to imag- of Decoy1.
ine they had 5000 Euros for investment. In this case they
only could choose among various bankbooks. Depending on • H4: Choice of Bankbook2 is increased by the presence
the products the subjects had to choose from, four groups of Decoy2.
4
• H5: Choice of Bankbook3 is increased by the presence Additionally, two asymmetrical dominated decoy items dA
of Decoy3. (decoy for A) and dB (decoy for B) were defined by choos-
ing the items with the second best interest rates of the two
Results and four year categories. The extreme products with a bind-
In this case users preferred high return rates over binding ing period of one and five years showing the highest interest
in years. Comparing the number of subjects choosing bank- rates in the respective categories were constituting the corre-
book1 in the control group and in the group Decoy 1 one sponding compromise decoys cA (decoy for A) and cB (decoy
can remark a non-significant increase by 1.5% (Fisher’s Ex- for B). Table 5 and Figure 7 are showing the resulting prod-
act Test, one-sided: p <.448, see Figure 6) [H3]. Comparing uct landscape of the experimental items. It has to be noted
the choice distribution of the control group and group De- that the design is not completely symmetric as the domi-
coy 2 the e↵ect was significant. The increase of bankbook2 nated items (dA and dB) are always inferior in the binding
choices in presence of decoy2 made up 12.1% (Fisher’s Exact dimension, such that dA constitutes a d3-decoy (see Figure
Test, one-sided: p <.001) [H4]. The decoy3 in the group De- 1) whereas dB constitutes a d2-decoy.
coy 3 increased the bankbook3 choices by 6.8% compared to
the control group (Fisher’s Exact Test, one-sided: p <.127) Item A B dA dB cA cB
[H5]. Interest
rate p.a. 3,00 3,77 2,75 3,60 2,25 4,00
(%)
Binding 2 4 2 4 1 5
(years)
Bank Deniz Auto Erste Direkt Direkt Direkt
Table 5: Attribute values of the experimental items
used in the di↵erent experimental groups.
Figure 6: Choice distribution for the ADE-task.
4. EXPERIMENT WITH A REAL-WORLD
SET OF FINANCIAL SERVICES Figure 7: Product landscape: two alternative items
Design with the corresponding decoy items.
The first step in order to come up with a realistic set of
items was to find a suitable product domain. The domain Grounding on the experimental super set in Table 5 and
of capital savings books was found to be perfect for our Figure 7, experimental sets were defined and categorised ac-
purposes because of the following reasons: cording to the decoy added to the core setting (i.e. only
the competing items A and B). The control (Control) set
• Savings books can be well described by two dimen- is consisting of only the competing items A and B. In the
sions, which is binding (i.e. the period in which it is decoy sets one out of four possible decoys (dA, dB, cA, cB)
not possible to withdraw the money) and interest rate was added, which should evoke the asymmetric dominance-
(p.a.). This o↵ers the possibility to stick to the simple or compromise e↵ect (ADE or CE) for the benefit of A or
two-dimensional item landscape. B, respectively.
• There is lots of comparable market data available. SetId Item 1 Item 2 Item 3 Decoy Type
0 A B Control
The experimental items for the di↵erent choice sets were 1 A B dA ADE - pro A
chosen on the basis of a products list given by Konsument.at, 2 A B dB ADE - pro B
3 A B cA CE - pro A
a well-known independent consumer information site.1 . Kon-
4 A B cB CE - pro B
sument.at listed capital savings books having a binding pe-
riod between one and five years. The products of the two- Table 6: Experimental item sets and type of decoy.
and four year categories having the highest interest rates
of that category were chosen as competing items A and B. The experiment was designed to be carried out online and
1 unsupervised. Subjects (students of University of Klagen-
Please note that the experiment was already carried out in
2009, such that the market data was up to date at this time. furt) were invited by email containing a link to the online
5
experiment to take part in the experiment. Figure 8 is show- In order to carve out the asymmetric influence of the de-
ing a screenshot of one experimental situation. Subjects coy on A and B, Table 8 lists only the choices of A or B, ne-
were asked to imagine to have 10000 Euros for investment glecting the decoy choices. Now it is revealed that except in
and to choose their favorite option out of a set of proposed group 3, where the relation between A and B kept almost the
items (i.e. the capital savings books). The subjects were same (i.e. H3 is not supported), the decoy pulled away more
assigned randomly to one of the defined settings. Further- choices from the competitor than from the target, i.e. dam-
more, the position of the presented items was random. aged the attraction of the target less than the attraction of
the competitor (i.e. H1, H2, H4 are supported). Hence, the
decoys rather caused an asymmetric detraction rather than
an asymmetric attraction. The reason, why there could not
be revealed a compromise e↵ect in group 3, is most probably
the distance between the decoy and the target (see Figure
7). The distance (i.e. cumulated attribute di↵erences) plays
a significant role for the strength of the decoy, such that the
bigger di↵erence is the less is the asymmetric influence of an
intended decoy.
Although the absolute choices of a target product are not
imperatively raised by a decoy item, there are nevertheless
two possibilities how bank institutes could benefit from de-
coy e↵ects. First, it is possible to shift the attraction within
a bank’s product assortments, as the bank’s reputation (i.e.
name) cannot have any influence (i.e. it is the same for all
products). For example, it would be possible to decrease
the attraction of products which show low marginal return
(i.e. competitor) or to increase the attraction of products
which show high marginal return. In this case, because of
the possibly many decoy choices, it would be crucial that
the decoy also shows a high marginal return rate in order to
improve the overall result.
The second possibility for exploitation addresses the pos-
Figure 8: Screenshot of the online experiment. The sibility for a bank itself being the target. In this case it
translations have been added post hoc. is more convenient to think about products as parts of the
bank’s product portfolio. When considering portfolios, the
Following the current theory, the control set should o↵er introduction of a decoy product could significantly take away
the most objective view on the competing items A and B as choices of the competitor banks portfolio for the sake of the
there are no decoy e↵ects, and thus should build the baseline. target bank’s portfolio. Thereby it does not matter which
With respect to this baseline, the following hypotheses were of the products in the portfolio benefits.
formulated:
SetId Decoy Type A B Decoy Total
0 Control 31 16 47
• H1: In setting 1, the asymmetric dominated decoy 66.0% 34.0% 100.0%
shifts attraction for the benefit of A (damages B). 1 ADE - pro A 31 9 9 49
63.3% 18.4% 18.4% 100.0%
• H2: In setting 2, the asymmetric dominated decoy 2 ADE - pro B 24 24 2 50
shifts attraction for the benefit of B (damages A). 48.0% 48.0% 4.0% 100.0%
3 CE - pro A 25 13 9 47
• H3: In setting 3, the compromise decoy shifts attrac- 53.2% 27.7% 19.1% 100.0%
tion for the benefit of A (damages B). 4 CE - pro B 20 16 18 54
37.0% 29.6% 33.3% 100.0%
• H4: In setting 4, the compromise decoy shifts attrac-
tion for the benefit of B (damages A). Table 7: Results of the experiment.
Results
Table 7 shows the experimental outcome for all five set-
tings. It becomes obvious that only in group 2 the decoy was 5. RELEVANCE FOR E-SALES SYSTEMS
able to lift the number of target choices. In the other groups In principle, decoy e↵ects occur in any system where com-
it seems that the choices of decoys were too many such that peting choice options are presented concurrently. Obviously,
the absolute number of choices of both, A and B, were de- this is the case for many e-sales systems like shop applica-
creased. For the groups 3 and 4, this is not surprising, as tions, recommender- and configurations systems, or many
non-dominated decoys (like a CE decoy) do not represent other online decision support systems. Although, depending
inferior options. The reason why the decoy in group 1 was on the application, there are various situations during the
chosen unexpectedly often must be the bank name. Whereas user sessions where cognitive biases like decoy e↵ects can
all other decoys were products from ’Denizbank’, the decoy play an important role, the most important phase for decoy
in group 1 was a product of ’Erste Bank’, which obviously e↵ects constitutes the product presentation phase. During
is a bank with better reputation. this phase purchase o↵ers (in shopping systems) or recom-
6
SetId Decoy Type A B Total future work is the implementation of a framework which
0 Control 31 16 47
allows to identify and control decision biases. In particular,
66.0% 34.0% 100.0%
1 ADE - pro A 31 9 40 we are working on a decoy filter for recommender systems
77.5% 22.5% 100.0% which is able to identify biased item sets and calculates a
2 ADE - pro B 24 24 48 set of items to be removed or added in order to objectify
50.0% 50.0% 100.0% the decisions. Specifically in the context of recommender
3 CE - pro A 25 13 38 systems this could lead to a big improvement in terms of
65.8% 34.2% 100.0%
recommendation accuracy and user trust.
4 CE - pro B 20 16 36
55.6% 44.4% 100.0%
Acknowledgement
Table 8: Results of the experiment, leaving out de- The work presented in the paper has been conducted within
coy choices. the scope of the XPLAIN-IT project which is financed by
the Privatstiftung Kaerntner Sparkasse.
mended items (in recommender systems) are typically pre-
sented concurrently and the user (consumer) finds himself in 7. REFERENCES
some sort of decision dilemma. Here, decoy e↵ects can man- [1] D. Ariely, T. Wallsten, Seeking subjective dominance in
ifest in suboptimal decision making as decoy e↵ects bias the multidimensional space: An exploration of the
perceived utility of the concurring options. This may fur- asymmetric dominance e↵ect, Organizational Behaviour
ther result in product purchases which are not optimal for and Human Decision Processes, 63(3), 223-232, 1995.
the consumer, the vendor, or both. [2] R. Burke, Knowledge-based recommender systems.
In the case of dialog-based systems (i.e. systems which Encyclopedia of Library and Information Systems,
gather user information by posing questions and proposing 69(32):180-200, 2000.
possible answers) decoy e↵ects can also influence the answers
[3] R. Burke, Hybrid recommender systems: Survey and
given by the users during the dialog. This can influence the
experiments. User Modeling and User-Adapted
accuracy and furthermore the time-efficiency of such sys-
Interaction, 12(4):331-370, 2002.
tems. For case case-based systems like tweaking-critiquing
[4] S. Callander and C. H.Wilson, Context-dependent
recommenders with multiple items to be criticized concur-
Voting, Quarterly Journal of Political Science, 1:
rently [5] this is obvious as any cycle basically constitutes a
227-254, 2006.
new product presentation phase.
Another aspect which is somewhat orthogonal to the bi- [5] L. Chen, P. Pu, Interaction design guidelines on
asing of decisions is the fact, that decoy e↵ects can have a critiquing-based recommender systems, User Modeling
positive e↵ect on the decision confidence [30]. This means and User-Adapted Interaction 19(3): 167-206, 2009.
that decoys manage to seemingly alleviate a decision situa- [6] A. Colman, B. Pulford, and F. Bolger. Asymmetric
tion such that users feel more confident about their decisions. dominance and phantom decoy e↵ects in games,
Altogether, the above mentioned aspects o↵er a big poten- Journal of Organizational Behavior and Human
tial for e-sales systems for optimizing the decision making Decision Processes 104(2007): 193-206.
process and also the quality of the taken decisions. [7] A. Felfernig, G. Friedrich, and L. Schmidt-Thieme.
Introduction to the IEEE Intelligent Systems Special
6. CONCLUSIONS AND FUTURE WORK Issue: Recommender Systems, 22(3):18-21, 2007
[8] Gigerenzer, G., Bounded rationality: Models of fast
In this paper we have presented the results of a series of
and frugal inference, Swiss Journal of Economics and
empirical studies that clearly show the impact of di↵erent
Statistics, 133(2/2), 201-218, 1997.
types of decoy e↵ects on the item selection behavior of a user
[9] Gigerenzer, G., Todd, P., ABC Research Group, Simple
in the context of financial services decision making. The ex-
heuristics that make us smart, New York/Oxford:
istence of decoy e↵ects has been shown for non-classical sce-
Oxford University Press, ISBN: 9780195121568, 1999.
narios with more than three items in the result set in order
to show the existence of decoy e↵ects for real world scenar- [10] J. L. Herlocker, J. A. Konstan, L. G. Terveen, and J.
ios. Therefore, we analyzed the existence of decoy e↵ects on T. Riedl, Evaluating Collaborative Filtering
the basis of the bankbook dataset provided by the Austrian Recommender Systems, ACM Transactions on
consumer advisory platform konsument.at. The results of Information Systems, 22(1):5-53, 2004.
our studies have a significant impact on the design of fu- [11] Huber, J., Payne, J. W., Puto, C., Adding
ture e-sales systems since it is obvious that item selection Asymmetrically Dominated Alternatives: Violations of
behavior is not based on a complete analysis of the set of of- Regularity and the Similarity Hypothesis, The Journal
fered or recommended items. Item selection is often subject of Consumer Research, Vol. 9(1), 90-98, 1982.
to the application of a set of simple heuristics which is the [12] Kahneman, D., Maps of bounded rationality:
reason for the observed decoy e↵ects. Taking into account psychology for behavioral economics, The American
these heuristics, and thus better understanding human de- Economic Review. 93(5), 1449-1475, 2003.
cision taking, can have positive e↵ects in terms of a higher [13] J. Konstan, B. Miller, J. Herlocker, L. Gordon, and J.
confidence in the set of presented items. Moreover, control- Riedl. GroupLens: Applying Collaborative Filtering to
ling such e↵ects also o↵ers the possibility of increasing the Usenet News, Communications of the ACM, 40(3):
probability of selection of certain items. 77-87, 1997.
Apart from the ongoing investigation of diverse decision [14] F.Y. Kuoa, T.H. Chub, M.H. Hsuc, H.S. Hsieha. An
biases in the context of e-sales systems, a main focus of our investigation of e↵ort-accuracy trade-o↵ and the impact
7
of self-efficacy on Web searching behaviors, Decision [32] D. Winterfeldt, W. Edwards, Decision Analysis and Be-
Support Systems, 37: 331-342, Elsevier, 2004. havioral Research, Cambridge University Press,Cambridge,
[15] R. J. Mooney, L. Roy, Content-based book England, 1986.
recommending using learning for text categorization,
5th ACM Conference on Digital Libraries, ACM Press,
2000.
[16] M. Ouyang, Does the Decoy E↵ect Exist in the
Marketplace? An Examination of the Compromise
E↵ect, Congress 2004 de l’Association des Sciences
Administrative du Canada, 2004.
[17] R. Paramesh, Independence of Irrelevant Alternatives,
Econometrica, 41(5):987-991, 1973.
[18] J.W. Payne, J.R. Bettman, and E.J. Johnson. The
Adaptive Decision Maker, Cambridge University Press,
Cambridge, England (1993)
[19] M. Pazzani and D. Billsus. 1997. Learning and
Revising User Profiles: The Identification of Interesting
Web Sites, Machine Learning. (27), 313-331, (1997).
[20] J. Quesada, N. Chater, P. Otto., and C. Gonzalez., An
explanation of decoy e↵ects without assuming
numerical attributes. 27th Annual Meeting of the
Cognitive Science Society. Chicago Lawrence Erlbaum
Associates, 2005.
[21] S. Ratneshwar, A. D. Shocker, D. W. Stewart, Toward
understanding the attraction e↵ect: the implications of
product stimulus meaningfulness and familiarity,
Journal of Consumer Research, 13:520-533, 1987.
[22] Rubinstein, A., Modeling Bounded Rationality, The
MIT Press,
Cambridge(Massachusetts)/London(England),
ISBN-10: 0262681005, 1998.
[23] C. Schmitt, D. Dengler, M. Bauer. The MAUT
Machine - an Adaptive Recommender System, ABIS,
2001.
[24] M. Schweizer, Kontrast- und Kompromisse↵ekt im
Recht am Beispiel der lebenslaenglichen Verwahrung,
Schweizerische Zeitschrift fur Strafrecht,
123(4):438-457, 2005.
[25] Simon, H. A., A behavioral model of rational choice,
The quarterly Journal of Economics, Vol. 69, 99-118,
1955.
[26] Simon, H. A., Theories of Bounded Rationality,
Decision and Organisation, Radner and Radner (Eds.),
1972.
[27] Simon, H. A., From Substantive to Procedural
Rationality, Method and Appraisal in Economics,
Latsis (Eds.), 1976.
[28] I. Simonson,A. Tversky, Choice in context: Tradeo↵
contrast and extremeness aversion, in: Journal of
Marketing Research (39), 281-292, 1992.
[29] Teppan, E. C., Recommendation beyond rationality,
Dissertation, University of Klagenfurt, 2010.
[30] Teppan, E., Felfernig, A., Impacts of Decoy Elements
on Result Set Evaluation in Knowledge-Based
Recommendation, International Journal of Advanced
Intelligence Paradigms (IJAIP), Vol. 1(3), 358-373,
2009.
[31] B. Wernerfelt, A Rational Reconstruction of the
Compromise E↵ect: Using Market Data to infer
Utilities, Journal of Consumer Research, 21:627-33,
1995.
8