Content uploaded by Alexander Felfernig
Author content
All content in this area was uploaded by Alexander Felfernig on Oct 22, 2014
Content may be subject to copyright.
Persuasive Recommendation: Serial Position Effects in
Knowledge-based Recommender Systems
A. Felfernig1, G. Friedrich1, B. Gula2, M. Hitz3, T. Kruggel1, G. Leitner3,
R. Melcher3, D. Riepan1, S. Strauss2, E. Teppan1, O. Vitouch2
1Intelligent Systems and Business Informatics
2Cognitive Psychology
3Interactive Systems
Klagenfurt University, Universitaetsstrasse 65-67, A-9020 Klagenfurt, Austria
{alexander.felfernig, gerhard.friedrich, bartosz.gula, thomas.kruggel, rudolf.melcher,
gerhard.leitner, daniela.riepan, sabine.strauss, erich.teppan, oliver.vitouch}@uni-klu.ac.at
Abstract. Recommender technologies are crucial for the effective support of
customers in online sales situations. The state-of-the-art research in
recommender systems is not aware of existing theories in the areas of cognitive
and decision psychology and thus lacks of deeper understanding of online
buying situations. In this paper we present results from user studies related to
serial position effects in human memory in the context of knowledge-based
recommender applications. We discuss serial position effects on the recall of
product descriptions as well as on the probability of product selection. Serial
position effects such as primacy and recency are major building blocks of
persuasive, next generation knowledge-based recommender systems.
Keywords: persuasive technologies, recommender systems, knowledge-based
recommendation, human memory, interactive selling.
1 Introduction
Recommender systems are among the most successful applications of Artificial
Intelligence technologies. The major purpose of recommender systems is to improve
the accessibility of complex and large product assortments for online customers.
There are basically three different types of recommendation approaches. One of the
most frequently used one is Collaborative Filtering [16, 29]. It implements the idea of
word-of-mouth promotion where a buying decision is predominantly influenced by
the opinions of friends and benchmarking reports. For instance, if two customers have
bought similar books in the past and have rated those books in a similar way,
positively rated books bought by only one of them, are recommended to the other
customer. Content-based Filtering [26] is an information filtering approach that
exploits item features a user has liked in the past to recommend new items. In contrast
to collaborative approaches, content-based filtering cannot provide serendipitous
recommendations. It recommends all items based on purchase information available
2
from the current user. Both approaches are based on long-term user profiles and do
not exploit deep knowledge about the product domain. Thus, they are excellent
techniques supporting recommendation processes for simple products such as movies,
compact discs or books. Compared to users purchasing simple products, those
purchasing complex products such as financial services or digital cameras are much
more in the need of information and in the need of intelligent interaction mechanisms
supporting the selection of appropriate items. Knowledge-based approaches [6,9]
make use of an explicit representation of product, marketing and sales knowledge.
Such deep knowledge allows (a) the recommendation of items which fulfil certain
quality requirements, (b) the explanation of recommended items, and (c) the support
of users in situations where no solution can be found. In contrast to word-of-mouth
promotion implemented by collaborative filtering, knowledge-based recommendation
implements explicit sales dialogs which support users in the item selection process. In
this paper we focus on knowledge-based recommender technologies that determine
recommendations on the basis of explicit sales dialogs where users are confronted
with questions related to their wishes and needs (preference elicitation phase – e.g.,
the tent recommender in Fig. 1). Elicited preferences are in turn used for calculating
recommendations for the current user. After the completion of a sales dialog, a
product comparison page is presented to the user which contains a set of alternative
items (see Fig. 1). The simplified tent recommender depicted in Fig. 1 has been used
as the basic stimulus/framework for user studies which are presented in the following
sections.
posed
questions
product
comparison
recommen-
ded items
Fig. 1: Example Recommender Application.
Knowledge-based recommender technologies have been successfully applied in
different commercial environments, for example the recommendation of financial
services [9] or restaurants [6]. A major reason for the successful deployment of those
technologies is that users do not only receive recommendations but additionally are
provided with a corresponding set of explanations as to why a certain item fits to the
wishes and needs of a user. Features such as explanations significantly improve the
trust of users regarding recommendations [9]. However, the development of
recommender applications is still rather focused on an existing set of technical
features. The effects of applying different theories about human memory in online
Serial Position Effects in Knowledge-based Recommendation 3
buying situations have not been analyzed up to now. In this paper we present results
of two empirical studies which investigate serial position effects [24] of human
memory in the context of recommendation sessions. Primacy and recency as a
specific form of serial position effects describe the phenomenon that information units
at the beginning and at the end of lists are more likely to be remembered than those in
the middle [10, 17]. Such effects may potentially occur in every situation where
information is presented in list format. In knowledge-based recommenders, there are
mainly three such listings: first, sequential product attribute questions in the dialog
phase. Second, the order of product attributes on the product comparison page, and
finally the order of products on the product comparison page.
In the relevant literature it has been argued that recommenders always persuade
when recommending [14, 20, 35]. This interpretation is based on the fact that
recommenders successfully support the effective identification of items which
otherwise would not have been found by the customer and consequently not been
purchased. We follow the definition of persuasion given in [10] where persuasion is
defined as the attempt of changing people’s attitudes or behaviours or both. Our
overall hypothesis is that serial position effects can be successfully exploited for
changing people’s attitudes in the context of online buying situations. In contrast to
[14, 20, 35] our approach actively exploits psychological theories for attaining
persuasion effects. For this purpose, knowledge-based recommenders can constitute
an ideal platform for installing persuasion technology. The deep understanding of
persuasion mechanisms offers the possibility of exploitation and control. COHAVE1
is an interdisciplinary research project at Klagenfurt University with the goal of
building a general framework for exploiting persuasive mechanisms in knowledge-
based recommendation. In this project, psychological theories from the areas of
memory phenomena and decision theory are investigated, implemented and evaluated.
The remainder of the paper is organized as follows: in Section 2 an overview of
related work is given and research questions are presented which are investigated in
the follow-up sections. Section 3 and Section 4 present results of two studies
investigating serial position effects in the domain of tents and digital cameras. The
paper is concluded with Section 5 where an outlook on future work is given.
2 Persuasive Effects in Preference Construction
Position effects in human memory are one of the oldest phenomena investigated in
experimental psychology [8, 19, 21, 24, 34]. Serial position effects are basic memory
phenomena first discussed in 1878 [24]. The effect has originally been found in short-
term memory tasks. It describes a specific order in the recall of a list of items, such as
meaningless syllables [8], numbers [24] or names of common objects [19] which
people had to learn by heart beforehand. In this context, recall accuracy of items from
a list shows two patterns: a) items from the beginning of the list (primacy) and b) the
items from the end of the list (recency) are better remembered than items from the
middle of the list [13, 24]. Mostly, primacy and recency effects have been explained
1 COHAVE is the acronym for Consumer Behavior and Decision Modeling for Recommender
Systems funded by the Austrian Research Fund (FFG-810996)., see cohave.ifit.uni-klu.ac.at.
4
as effects of the dual store account model of human memory [1], but there is also
evidence for a serial position effect related to long-term episodic memory [23].
Order effects in persuasion and ‘the motivation to think’ have been discussed for
example in [27]. It could be shown that under chunked conditions, participants who
were highly motivated to think were more susceptible to primacy and recency effects
than those low in motivation to think. There are numbers of studies dealing with both
short- and long-term episodic memory tasks. The outcome of studies of long-term
serial position effects [2, 17, 28, 30] using serial order reconstruction tasks show clear
recency effects. [23] shows a corresponding effect in semantic memory tasks using
verse hymns as stimuli, resulting in the first unequivocal demonstration of serial
position effects in semantic memory. In contrast to most other work mentioned above
we use meaningful product-features (questions) as stimuli which are used as
information units in knowledge-based recommender systems. Information in
knowledge-based recommender systems is usually presented in the form of ordered
lists of questions, product attributes, and recommended products. Unlike meaningless
material, this kind of information requires a higher level of semantic processing.. The
studies presented in this paper deal with primacy and recency effects in a semantic
memory task and focus on the dialog and the product selection phase.
Research on consumer buying decision making argues that preferences are rather
constructed spontaneously [3, 5, 25] than being stable. Following this interpretation,
studies have recently shown several psychological phenomena that affect these short-
term processes of preference construction. Through feature-based priming for
instance, the background of an e-commerce site can guide the attention of customers
towards specific product attributes [22]. The attention can also be influenced by the
inclusion or exclusion of attributes in the dialog of a recommender system [15]. Both
mechanisms contribute to the construction of consumer preferences and to the
consideration of product attributes that otherwise may have been omitted. Taking into
account these mechanisms can create a new possibility for product suppliers on e-
commerce sites to emphasize on those product attributes with which they can
outperform their competitors.
The major goal of this paper is to investigate to which extent serial position effects
occur in the context of knowledge-based recommenders. Once serial position effects
have been proven to work for such dialog systems, mechanisms for exploiting these
effects can be implemented in knowledge-based recommender applications. The
primacy and recency effect would thus influence the design of recommendation
dialogs in terms of question ordering as well as the ordering of the product features.
We assume that a supplier who tries to ‘positively convince’ (persuade) a customer of
the quality of certain products should present the best attributes of her products at the
beginning and at the end of product descriptions or result pages of a recommender-
application. We examined our assumptions in two studies. Study 1 addressed the
general question whether serial position effects occur for the recall of product
attributes in the dialog phase of a recommender. In Study 2 we investigated whether
serial position effects from the dialog phase directly influence product selection. In
this context, we focused on answering the following research questions:
Serial Position Effects in Knowledge-based Recommendation 5
o Q1: Do serial position effects exist for sentences and product feature
descriptions?
o Q2: Do serial position effects occur across different product domains?
o Q3: Do serial position effects influence the importance of attributes in a
purchase situation?
o Q4: Do serial position effects in the dialog of a recommender influence product
choices of customers?
o Q5: Are product choices influenced by the order of attributes or products on a
product comparison page of a recommender?
3 Serial Position Effects in the Recall of Product Descriptions
The goal of the study 1 was to investigate serial position effects in the recall of
product descriptions related to tents and digital cameras. In this study, 14 product
attributes of tents as well as of digital cameras were collected. For each product
attribute a corresponding explanatory sentence has been formulated (e.g., ‘with a
waterproof tent you can camp on rainy days’ or ‘the lowest capacity of memory cards
for digital cameras is 16 megabytes’). Such explanatory sentences have been
integrated in a MS PowerPoint presentation with one sentence per slide. Each slide
has been presented for 15 seconds. First, participants had to read each explanatory
sentence. Subsequently the participants had to recall as many attributes from the list
as possible (after viewing the whole slideshow). Immediately after the recall task,
participants were asked to rate the importance of each attribute they remembered
would have in a real purchase decision as well as to estimate the overall familiarity of
an average consumer with an attribute on a 5-point Likert scale.
In order to design orthogonal attribute orders, an a priori expert rating for the
expected overall familiarity of customers with product attributes has been performed.
Based on this rating, two different attribute sequences (lists) have been implemented
for each product domain by categorizing the attributes as familiar salient and
unfamiliar salient. In the familiar salient-list the most familiar attributes were
positioned in the beginning and end of the lists while the less familiar attributes were
put in the middle. In the unfamiliar salient-list the less familiar attributes were
presented in the beginning and end of the lists. The experiment was conducted with
four different groups of subjects. In each group participants were confronted with one
list version for digital cameras and one list version for tents (see Table 1).
group attribute sequence 1 attribute sequence 2
1 digi_familiar_salient tents_ familiar_salient
2 tents_unfamiliar_salient digi_ unfamiliar_salient
3 digi_ unfamiliar_salient tents_ familiar_salient
4 tents_ unfamiliar_salient digi_ familiar_salient
Table 1: Groups and Attribute Sequences.
6
N = 72 students of the Klagenfurt University (36.1 % female) with a mean age of 23.3
years (SD = 5.1) were tested in group sessions. Out of the 14 product attributes
subjects recalled 8.2 attributes of tents (SD = 4.0) and 8.0 attributes of digital cameras
(SD = 3.38). This difference is not significant.
Results for tents. For the analysis, attributes were combined into pairs according to
their position within each list. The results of a computed two-factorial ANOVA show
that the position of an attribute pair has a clear effect on the frequency of recall (F(6,
70) = 5.75, p < .001, η2 = .08, see Fig. 2). The list-version had no influence on the
frequency of recall (p = .34). Descriptively, the slightly incremented recall for middle
attribute pairs (3-5) in the unfamiliar salient list reflects the fact, that in this list more
familiar attributes were presented in the middle.
Fig. 2. Relative frequencies of recall for consecutive attribute pairs of tents (1-7). The
continuous line corresponds to the results for the unfamiliar salient- and the dashed line to the
familiar salient-list. The bars represent the standard errors in all figures.
The probability of recalling attributes from the first pair was .8 and the last pair .72.
Combined over both lists we first tested the difference in recall between the first item
pair and each of the remaining six pairs, and second, between the last pair and all the
other pairs. The investigation of these specific contrasts results in a clear pattern: the
probability to recall either the first (primacy) or last (recency) pair was significantly
higher than the probability to recall any of the attribute pairs in the middle of the lists
(five F-tests, all p < .01). At the same time the recall performance for the first and last
pairs did not differ significantly. Combining the attributes in the middle into one
group shows an even more pronounced position effect (F(2, 70) = 13.28, p < .001, η2
= .16). The self reported knowledge about tents was coded into a dichotomous
variable using a median split and has been included in the analysis. Subjects reporting
higher knowledge were tending to recall more attributes. However, at least for tents
serial position effects occurred independently of the self-reported product domain
knowledge.
Serial Position Effects in Knowledge-based Recommendation 7
Results for digital cameras. We found a significant interaction between attribute
position and attribute familiarity (F(6, 70) = 6.05, p < .001). Both serial position
effects (primacy and recency) can only be found in the familiar salient-list which
contained the more familiar attributes at the beginning and at the end (see Fig. 3). The
pattern of results for specific contrasts is less clear than for tents: first and last
attribute pairs were recalled significantly more often than the three pairs in the middle
of the list but the differences to the second and last but one pair were not significant.
Because there are no guidelines on how many items are to be involved in primacy and
recency effects, the choice of pairs is arbitrary. Especially, for the pattern of results
shown in Fig. 3, it seems more plausible to assume that all four attributes presented at
the beginning of the list contributed to a primacy effect. For the unfamiliar salient-list
it is noticeable that if no position effects occurred and only attribute familiarity
influenced recall performance, the expected line in Fig. 3 should be inversely u-
shaped. In this list, the most familiar attributes which were in the middle of the list,
would be recalled more often than the less familiar attributes at the beginning and end
of the list, which is not the case. A possible explanation would be that primacy and
recency actually occurred in the unfamiliar salient-list and resulted in an improved
recall performance on unfamiliar attributes.
Fig. 3. Relative frequencies of recall for consecutive attribute pairs of digital cameras (1-7).
The continuous line corresponds to the results for the unfamiliar salient- and the dashed line to
the familiar salient-list.
Summarizing, serial position effects do exist for descriptions of product features
presented subsequently (Q1). However, the effect was not as domain-independent as
assumed manifesting itself less clearly in the domain of digital cameras (Q2). Also,
the self reported domain knowledge did not suppress the effect. More knowledgeable
participants also remembered attribute descriptions from the beginning and end of the
lists more often then attributes in the middle.
8
Participants rated resolution and zoom of digital cameras to be the most important
attributes in a real purchase situation and waterproofness and insect protection as the
most important attributes of tents. In order to assess whether the position of attributes
influenced the importance ratings (Q3), a two-factorial ANOVA for list-version and
position was computed with three positions (beginning: first pair, middle: all five
pairs in the middle and end: last pair). The importance ratings for digital cameras
showed an interaction between list-version and position (F(2, 56) = 21.26, p < .001).
The pattern does not seem to resemble an influence of serial position effects on
importance ratings because more familiar attributes at the beginning and end were
rated significantly more important than the attributes in the middle for the familiar
salient-list and vice versa for the unfamiliar salient-list. This result shows that familiar
attributes are rated as important. However, importance ratings of single attributes did
differ according to our expectation depending on their position. For example
additional lenses were rated significantly more important in the unfamiliar salient-list
where this attribute was presented first than in the familiar salient-list where it was in
the middle (t(34) = -1.71; p = .04). Importance ratings for attributes of tents varied
depending on their position in the list. Attribute pairs at the beginning were rated
more important than those in the middle (F(1, 76) = 13.92; p < .001; η2 = .16) and also
attribute pairs at the end were rated more important than those in the middle (F(1, 76)
= 4.85; p = .03; η2 = .06). The effect is larger for primacy than recency. This result
implies that at least for tents the sequential order of product descriptions influences
importance ratings and thus may influence actual product purchases. Among others,
this question is pursued in the following study performed with an actual recommender
(tent recommender application).
4 The Influence of Serial Position Effects on Product Choice
In order to test whether positions of product attributes in the dialog and product
comparison page influence product choice we have constructed six versions of a tent
recommender with 10 attributes. In a two-factorial ANOVA we first varied three
different attribute orders in the dialog (random order, fixed order 1 and fixed order 2)
and combined it with two different orders of attributes on the product comparison
page. In both orders of attributes in product comparison the first four attributes listed
were the same as the first and last two in the corresponding dialog (see Table 2).
To be able to compare product choices over all six versions we presented to each
participant the same set of four tents on the comparison page. The four tents were
defined by using the attribute importance ratings from study 1 (see Section 3). The
multi-attribute utility value was about the same for each tent. Two of the tents were
defined as ‘target products’, because they outperformed all others when judged on the
first and last two attributes from the corresponding dialog only. If serial position
effects from the dialog influence the perceived importance of attributes (as shown in
study 1), participants should choose the target product more often when interacting
with the recommender with fixed order 1 in the dialog compared to any other order.
The order of products on the product comparison page was random for each
participant. The task of the participants was first to choose the tent they would buy
Serial Position Effects in Knowledge-based Recommendation 9
most likely in a real purchase situation and second, to rank all four tents’
attractiveness. Participants were recruited from students of the Klagenfurt University.
The possibility of winning 1 x €100 and 2 x €50 has been offered. Participants were
randomly assigned to one of the six versions of a tent recommender. Finally, 650
valid sessions could be extracted from the log files. Mean age of participants was 25.3
years (SD = 6.48), 63 % of them were female. The median time to complete the dialog
was ~2.5 minutes and it took ~1.8 minutes to choose and rate products.
Comparison
Ordering 1
Dialog
Ordering 1
Dialog
Ordering 2
Comparison
Ordering 2
1 1 6 6 1 -waterproof
10 2 7 5 2 -insect protection
2 3 8 7 3 -air ventilation
9 4 9 4 4 -installation time
3 5 10 8
5 -roof of entrance
8 6 1 3
6 -weight
4 7 2 9 7 -extreme temperatures
7 8 3 2 8 -extreme crack resistance
5 9 4 10 9 -aerodynamic
6 10 5 1 10 -window
Table 2: Dialog and Product Comparison Orderings.
Results. Across all six recommender versions target product 1 was preferred more
than any of the three other products χ2(3, N = 650) = 636.54; p < .001). It is noticeable
that tent 1 outperformed all other tents in the set on two attributes rated as the most
important ones. However, at the same time it showed worse quality on six attributes
compared to one of the other tents. Taken together with the fact that products were
generated with similar multi-attribute utility based on ‘real’ importance ratings
(derived from study 1) this result suggests that participants based their choice only on
a few important attributes rather than using all available information to decide.
Support of this interpretation may be found in articles suggesting heuristic decision
models like the lexicographic strategy or elimination by aspects [32, 33].
A two-factorial ANOVA with kind of dialog (three levels: random dialog, dialog 1
and dialog 2) and kind of product comparison (two levels: comparison 1 and
comparison 2) was computed to determine effects on the relative frequency of
booking the target product. Opposed to our expectations (Q4 in Section 2), the
frequency of booking the target product was not affected by the order of attributes in
the dialog (F(2, 643) = .44; p = .65) but by the order of attributes on the product
comparison page (F(1, 643) = 9.76; p = .002). 74 % of subjects interacting with
product comparison 1 chose the target product but only 62 % of subjects interacting
with product comparison 2 (see Fig. 4). The choice of the target product was not
biased by subjective domain knowledge.
10
Fig. 4. Frequencies of product choice: tent 1 (black) vs. all other tents: (white).
To determine the relative impact of (a) kind of product, (b) its position on the
comparison page, (c) the question order in the recommender dialog, and (d) the
attribute order in the product comparison page on the choice behaviour of
participants, a four-way frequency analysis was computed. The hierarchical log linear
model describing the data best consists of two two-way interactions (position x
product and attribute order x product) with likelihood ratio χ2(76, N = 650) = 67.67; p
= .74. Especially the attribute order x kind of product interaction shows that
depending on the kind of attribute order each tent was chosen more or less often than
expected (see Q5 in Section 2). While the target product was chosen more often than
expected in product comparison 1 all other products were chosen more often than
expected in product comparison 2. The second interaction (product position x kind of
product) results from the fact that the target product was chosen more often than
expected when it was presented as the first or last of all four tents, while there was an
inverse trend for all other products. This implies that the order of products on product
comparison pages has an influence on product choice. Summarizing, the order of
attributes has an impact on product choice. Results of study 1 imply that attribute
order in the dialog has an impact on the perceived attribute importance. Contrary to
this result, no such impact of the dialog on product choices could be found in study 2
(a further clarification is needed in this context).
5 Conclusions and Future Work
The studies presented in this paper show that in the line of feature-based priming
and inclusion effects serial position effects are another interesting cognitive
phenomenon that can play a crucial role in the design of product comparison pages in
recommender systems. This result generalizes beyond the dialog of knowledge-based
recommenders and can be applied to a wide variety of product and service
Serial Position Effects in Knowledge-based Recommendation 11
descriptions ranging from product fact sheets, package leaflets, motivational
campaigns for the participation in health promotion or political engagement programs.
Based on the results reported in this paper, several challenges in the design of
knowledge-based recommender applications emerge. It seems that long attribute lists
are not necessary for users’ decisions. Furthermore, algorithms are needed that
provide as little information as necessary and as much as needed to not reduce a
users’ trust in the recommender application. In relation to the latter arguments, it does
matter how attributes are ordered on product comparison pages and a corresponding
recommendation to developers of recommender applications can be made to actively
take into account serial position effects when designing result (product comparison)
pages for recommenders.
References
1. R. C. Atkinson and R. M. Shiffrin. Human Memory: A proposed system and its
control processes. In K. W. Spence and J. T. Spence (Eds.), The psychology of
learning and motivation: Advances in research and theory (pp. 89-195). New York:
Academic Press, 1968.
2. A. D. Baddeley and G. Hitch. Recency re-examined. I S. Dornic (Ed.), Attentional
and performance VI (pp. 647-667). Hillsdale. New York: Erlbaum, 1977.
3. S. Bellman, E. J. Johnson, G. L. Lohse and N. Mandel. Designing Marketplaces of
the Artificial with Consumers in Mind: Four Approaches to Understanding Consumer
Behavior in Electronic Environments. Jrnl. of Interactive Marketing, 20: 21-33, 2006.
4. D. Berdichevsky and E. Neuenschwander. Towards an ethics of persuasive
technology, Communications of the ACM, 42 (5), 1999.
5. J. R. Bettman, M. F. Luce and J. Payne. Constructive Consumer Choice Processes.
Journal of Consumer Research, 25: 187-217, 1998.
6. R. Burke. Knowledge-based Recommender Systems. Encyclopedia of Library &
Information Systems, 69 (32), 2000.
7. R. Burke. Hybrid Recommender Systems: Survery and Experiments, User Modeling
and User-Adapted Interaction 12:331–370, 2002.
8. H. Ebbinghaus. Memory: A contribution to experimental psychology. New York:
Columbia University, Teachers College, 1913.
9. A. Felfernig and B. Gula. An Empirical Study on Consumer Behavior in the
Interaction with Knowledge-based Recommender Applications, in IEEE Joint
Conference on E-Commerce Technology (CEC06) and Enterprise Computing, E-
Commerce and E-Services (EEE06). San Francisco, California: IEEE Computer
Society, pp. 288–296, 2006.
10. B.J. Fogg. Persuasive computers: Perspectives and research directions, Proceedings
of the SIGCHI conference on Human factors in computing systems CHI '98, 1998.
11. B.J. Fogg. Persuasive Technology, Morgan Kaufmann Publishers, 2003.
12. R. Gasser, D. Brodbeck, M. Degen, J. Luthiger, R. Wyss and S. Reichlin.
Persuasiveness of Mobile Lifestyle Coaching Application Using Social Facilitation,
Persuasive 2006, pp. 27-38, 2006.
13. F. B. Gershberg and A. P. Shimamura. Serial position effects in implicit and explicit
tests of memory. Journal of Experimental Psychology: Learning, Memory, and
Cognition, 20: 1370-1378, 1994.
14. U. Gretzel, D. R. Fersenmaier. Persuation in Recommender systems, International
Journal of Electronic Commerce, 7(2): 81-100, 2007.
12
15. G. Häubl and K. B. Murray. Preference Construction and Persistence in Digital
Marketplaces: The Role of Electronic Recommendation Agents. Journal of
Consumer Psychology, 13: 75-91, 2003.
16. J. L. Herlocker, J. A. Konstan, L. G. Terveen, and J. T. Riedl, Evaluating
Collaborative Filtering Recommender Systems, ACM Trans. on Information Systems,
22(1): 5–53, 2004.
17. G. J. Hitch and J. Ferguson. Prospective memory for future intentions: Some
comparisons with memory for past events. European Journal of Cognitive
Psychology, 3: 285-295, 1991.
18. P. King, and J. Tester. The landscape of persuasion technologies, Communications of
the ACM, 42 (5): 31-38, 1999.
19. E. A. Kirckpatrick. An experimental study of memory. Psychological Review, 1: 602-
609, 1894.
20. S. Komiak and I. Benbasat. Comparing Persuasiveness of Different Recommendation
Agents as Customer Decision Support Systems in Electronic Commerce,
International Conference on Decision Support Systems (DSS 2004), Prato, Tuscany,
2004.
21. K.S. Lashley, K.S. The problem of serial order in behavior. In L. A. Jeffress (Ed.),
Cerebral mechanisms in behaviour (pp. 112-136). New York: Wiley, 1951.
22. N. Mandel and E.J. Johnson. When Web pages influence choice: Effects of visual
primes on novices. Journal of Consumer Research, 29: 235–245, 2002.
23. E. Maylor. Serial position effects in semantic memory: reconstructing the order of
verses of hymns. Psychonomic Bulletin & Review, 9: 816-820, 2002.
24. F. E. Nipher. On the distribution of errors in numbers written from memory.
Transactions of the Academy of Science of St. Louis, 3, CCX-CCXI, 1878.
25. J. W. Payne, J. R. Bettman and E. J. Johnson. The Adaptive Decision Maker. New
York: Cambridge University Press, 1993.
26. M. Pazzani and D. Billsus. Learning and Revising User Profiles: The Identification of
Interesting Web Sites. Machine Learning, 27: 313-331, 1997.
27. R. E. Petty, L. T. Zakary, C. Hawkins and D. T. Wegener. Motivation to think and
order effects in persuasion: the moderating role of chunking. Personality and Social
Psychology Bulletin, 27 : 332-344, 2001.
28. A. C. Pinto and A. D. Baddeley. Where did you last park your car? Analysis of
naturalistic long-term recency effect. European Journal of Cognitive Psychology, 3:
297-313, 1991.
29. B. Sarwar, G. Karypis, J. A. Konstan, and J. T. Riedl. Item-based collaborative
filtering recommendation algorithms. In 10th Int. World Wide Web Conf., pages 285-
295, 2001.
30. J. R. Sehulster. Content and temporal structure of autobiographical knowledge:
Remembering twenty-five seasons of the Metropolitan Opera. Memory & Cognition,
17. 590-606, 1989.
31. Y.-Y. Shih and D.-R. Liu. Hybrid recommendation approaches: collaborative
filtering via valuable content information, In 38th Hawaii International Conference
on System Sciences (HICSS’05), page 217b, Big Island, Hawaii, 2005.
32. A. Tversky. Intransitivity of preferences. Psychological Review, 76: 327-352, 1969.
33. A. Tversky. Elimination by aspects: A theory of choice. Psychological Review, 79:
281-299, 1972.
34. R. K. Young. Serial Learning. In T. R. Dixon and D. L. Horton (Eds.), Verbal
behaviour and behaviour theory (pp.122-148). Engelwood Cliffs, NJ: Prentice Hall,
1968.
35. M. Zanker, M. Bricman, S. Gordea, D. Jannach and M. Jessenitschnig. Persuasive
online selling in quality & taste domains, EC-WEB 2006, pp. 51-60, 2006.
