the individual to perceive the content and value of the recommendations in a certain way. For
example, a user may feel more emotionally attached to (and hence more influenced by)
information that comes from one individual compared with information generated by a machine.
To test this we compared both the product network (C1.1) and the dual network (C1.2)
conditions of study 1 to five additional conditions. First, we offered the user the dual network
recommendation (similar to C1.2), but changed the label of the user-generated links from
"People who liked this also liked" to "Videos liked by friends of this video author" (C4.1).
Second, we offered the user the dual network recommendation (similar to C1.2), but we
switched the labeling of the two groups of recommendations: links from the product network
were marked as "People who liked this also liked", and links based on the user-generated links
were labeled as "Related Videos" (C4.2). Third, we offered the user the dual network
recommendation (similar to C1.2), but all recommendations were provided under the same label
("Related Videos") (C4.3) or with no labels at all (C4.4). Fifth, participants were exposed to 10
recommendations based only on the YouTube product network (similar to C1.1), all under the
label of "People who like this also liked" instead of "Related Videos" (C4.5). Thus, the first four
conditions, C4.1-C4.4, offered the same recommendations as the dual network condition (C1.2)
described in study 1, but disguised under different labels. The fifth condition (C4.5) offered the
same recommendations as the product network condition described in study 1 (C1.1).
We carried out the study according to the procedure described in study 1, with 500
participants (48.5% female, mean age 28.3), each assigned randomly to one of the new
conditions. Using the same criteria as in study 1, we removed from the data 7 participants.
As in study 1, we compared the hazard ratios of the different conditions. The results show
that participants who were exposed to dual network recommendations (C1.2, C4.1, C4.2, C4.3
and C4.4) found “good” content more quickly than did those who were exposed to the product
network recommendations only (C1.1 and C4.5), no matter how those recommendations were
labeled (hazard ratios for the dual network conditions: HR
=1.296 and HR
=1.38; the hazard ratio of the product network labeled
as the dual network is HR
=0.957; all are compared to the product network condition HR
as a baseline; p < 0.05).
We also performed a contrast analysis (Rosenthal and Rosnow 1985), which revealed a
significant difference (F = 5.98, p<0.01) in overall satisfaction in favor of participants who used
the dual network (C1.2 and C4.1-C4.4), compared with participants exposed to the product
network structure only (C1.1 and C4.5).
Additionally, we find significant differences in the AverageRating and the
HighRankingScore variables (F = 4.26, p < 0.01 and F = 3.55, p < 0.01, respectively) when
contrasting participants in the dual network condition with those in the product network
condition. There were no significant differences for those measurements within participants in
the different dual network conditions (C1.2 and C4.1-C4.4) and no significant difference among
participants in the product network conditions (C1.3 and C4.5).
These findings show that participants exposed to dual network recommendations performed
better in finding "good" content than did those who were exposed to the product network
recommendations, regardless of the labeling of the recommendations implying that the mere
labeling has a marginal effect, if any.
THE EFFECT OF DYNAMIC STRUCTURE CHANGES
Unlike stores, electronic commerce sites (e.g., Amazon) can respond to consumers’
searches in real time and change displays dynamically to direct their exploration. Additionally,
as observed by Hui, Bradlow and Fader (2009), as the amount of time a consumer spends in a
store increases, the consumer’s tendency to continue exploring the store decreases. A possible
solution to this problem may be to dynamically change the “store” design throughout the
consumption process. Thus, rather than present consumers with a static view of the product
network, e-commerce sites can dynamically incorporate user-generated links to improve the
efficiency of the search process and to retain customers’ interest. When such a dynamic strategy
C# Excel - Excel Conversion & Rendering in C#.NET
independently, without using other external third-party dependencies like Adobe Acrobat. Excel to PDF Conversion. This page will tell you how to use XDoc.Excel adding a page to a pdf in preview; add contents page to pdf
is used, the question becomes, is there a point during the exploration process at which the
consumer might benefit more from the introduction of user-generated links?
The following studies were designed to provide initial answers to this question. In our
studies, each participant was exposed to different structures at different times during the
navigation process. The switch to a new network structure was based either on the phase of the
exploration process the user was in (studies 5 and 6) or on her satisfaction during exploration
Study 5: Structure Change as a Function of Exploration Time
In this study, we explore the conjecture that the dual network structure will be most
beneficial to consumers who have been browsing the website for a while and may need to revive
the search with more diversity. This conjecture is based on the belief that in a continuous
process of exploration, product network recommendation mechanisms tend to narrow down the
navigation possibilities to objects that are very similar to one another. This is also supported by
our results with respect to the high level of assortative mixing mentioned above. We therefore
created a new condition (C5.1) using a dynamic dual network and compared it to the previously
mentioned static conditions (product network, dual network). Under this new condition
consumers were presented with recommendations based on the product network structure and
after a specified amount of time were switched to the dual-network structure.
Method. In this study 300 participants (49% female; mean age 28.5) were each assigned
randomly to one of the following conditions. One-third of the participants were presented with
the dynamic dual network condition (C5.1), with the switch from product network to dual
network recommendations occurring after 250 seconds. The rest of the participants were
presented with the static product network (C1.1) and static dual network (C1.2) conditions. The
experimental procedure was the same as that described in study 1. Using the same criteria as in
study 1 we removed from the data 11 participants.
As in study 4, to control for a possible labeling effect and to avoid a visible change in the
appearance of the website in the middle of the study, we grouped all types of recommendations
under the same label ("Related Videos").
Analysis and Results. Given that this was a within-subject design, the hazard model
estimated in the previous section was no longer appropriate. Also, we could not simply compare
the ratings before and after the change in network structure, because our previous results
demonstrated a continuous decrease in users’ rankings over time regardless of the types of
recommendations presented. Hence, in order to replicate the previous results (dual network
effect) we needed to compare between and across subjects, using a difference-in-differences
model (Bertrand et al. 2004). In our context, the control group was the group presented with the
product network recommendations only, with no dynamic switch (i.e. C1.1). We estimated the
is the average rating in period t for participant i; Treatment
is a treatment
dummy variable, with T
= 1 representing exposure to the dynamic condition; Period
is a period
dummy variable, where Period
= 0 indicates the time period in the study before treatment (i.e.,
the time period before the switch to the dual network; t ≤ 250 seconds) and Period
indicates the time period after treatment (t > 250 seconds); and Treatment
interaction term and represents the actual treatment variable. Table 4 presents the estimates of
(Insert Table 4 about here)
The coefficient for Treatment is not significant, indicating that there are no significant
differences between the groups (treated and untreated) A priory. Similarly, the coefficient for
Period is not significant, indicating that there are no significant differences within subjects who
are not treated. The interaction coefficient is positive and significant, thus indicating that users
under the dynamic condition demonstrated a significant increase in rating after the switch to the
(Insert Table 5 about here)
As shown in Table 5, users of the dynamic dual network design also had higher satisfaction
than did users of the static product network design (p < 0.01) or the static dual network design
(p < 0.01). Looking at the AverageRating and the HighRankingScore, which are measures that
represent the entire process, we find that participants in condition C5.1 scored significantly
higher than did those in the static product network condition (p < 0.05). However, we find no
significant difference between participants of this time-dependent dynamic dual network
condition (C5.1) and those of the static dual network condition (C1.2).
Study 6: Testing Whether the Effect Is Driven by the Content of the Generated Links or Whether
This Is an Artificial Effect Caused by Mere Change.
It is possible that the effects observed in study 5 are merely the results of diversity: users
may simply enjoy the opportunity to view videos that seem different from the videos previously
offered. Thus, the results are not necessarily related to the content of the links in the dynamic
condition. Therefore, for robustness, and to rule out this alternative explanation, we carried out
an additional study, creating a new dynamic structure based on the randomized links.
Method. Under this condition (C6.1), 100 participants (50% female, mean age 27.6) were
presented with recommendations based on the product network structure and after a fixed
amount of time were switched to the randomized links structure (2 participants were excluded
based on the criteria described above). This is an important control, because if consumers only
seek diversity, a randomized links condition might be able to provide the same benefits as the
dual network. Again, to control for a possible labeling effect and to avoid a visible change in the
appearance of the website in the middle of the study, we grouped all recommendations under the
same label ("Related Videos").
Analysis and Results. We estimated a difference-in-differences model similar to the one in
study 5, with the treatment now being the switch to the randomized links. The results are
presented in Table 6 and show that in contrast to the time-dependent dynamic dual network
condition (C5.1), switching from the product network to the randomized links condition had a
significant negative effect, indicating, again, that the randomized links cannot mimic the
positive effect of the user-generated links and the dual network role.
(Insert Table 6 about here)
We also compared participants’ overall satisfaction, AverageRating, and HighRankingScore
(Table 5 above), and found that these measures were significantly lower among participants in
the dynamic randomized links condition (C6.1) than among those in C5.1, the time-dependent
dynamic dual network condition (p < 0.01), or in C1.2, the static dual network condition (p <
Study 7: Structure Change as a Response to Participants' Discontentedness
Our findings above show that switching from a product network to a dual network after a
certain period of time in the exploration process has a positive effect. In this study, we focus on
a dynamic environment in which the site structure changes in response to the consumer’s
reactions. Like study 5, this approach is based on the belief that during a continuous process of
exploration, product network recommendation mechanisms tend to narrow down the navigation
possibilities to very similar objects. When a user dislikes what she is watching, offering similar
objects will not be of great help, and it may therefore be time to point her to different areas in
the product network. Hence, introducing a dual network when the user is dissatisfied might
improve the exploration process.
Method. In this study, the dynamic structure was designed as follows. When a user began
the exploration process, the recommendations he or she received were based on a predefined
network structure. When (and if) the user ranked two consecutive videos with a low rating (<3),
the structure of the site changed so that the user was exposed to a different network structure.
We randomly assigned 300 participants (51% female, mean age 28.9) to the following
dynamic conditions: (1) starting from the product network and switching to the dual-network
structure for the rest of the exploration time (C7.1); (2) starting from the product network and
switching to the randomized links together with the product network structure for the rest of the
exploration time (C7.2); (3) starting from the dual network and switching to the product network
structure for the rest of the exploration time (C7.3). Since the change of structure was induced
by the participant’s activities, it was possible that a participant would not be exposed to a change
in structure at all. Nine participants were removed based on the criteria described above.
Analysis and Results. We found that nearly 80% of the participants experienced a sequence
of videos that did not match their preferences and were therefore exposed to a change of
structure during their exploration.
Since the time of the switch varied across participants, we could not use the difference-in-
differences model estimated above to analyze the results of this study and were therefore unable
to control for potential across-subject differences.
(Insert Table 7 about here)
Nevertheless, we were able to compare within subjects using paired sample t-tests. For each
participant we compared the average ranking before the switch with the average ranking after
the switch. As shown in Table 7, participants under condition C7.1 (dynamic dual network by
ranking, that is, shift from product to dual network structure) experienced the largest jump in
average rating. Among these participants, the switch increased video ratings by 17.5% on
average. In conditions C7.2 and C7.3, we did not find a significant difference between
consumers’ average ratings before versus after the switch. These findings suggest that a change
from the product network to the dual network as a response to the user's indication of
dissatisfaction may lead to an increase of her ranking of the products she consumes. We may
further conclude that the increase in the rating is not due to the effect of switching between
different structures but is influenced by the actual structure presented to the participant.
Documents you may be interested
Documents you may be interested