Docstoc

Contextual Multi-Dimensional Browse

Document Sample
Contextual Multi-Dimensional Browse Powered By Docstoc
					                              Contextual Multi-Dimensional Browse




ABSTRACT                                                                                                                                           Root

                                                                                                                                            p1                     p 12
Browsing, which usually alternates with search in informa-
tion seeking on the web, highly depends on information orga-
                                                                                                                                 a1                                              a2
nization of the browsed web site. An emerging web content
categorization is multi-faceted, which does not impose any
rigid structure, and allows users to slice and dice the infor-                                                    b1                        b2                       b1                   b2
mation space from any chosen “dimensions” they wish to
browse. However, owing to the limited capacity of human                                            c1        c2             c3         c1         c2          c2          c3      c1      c2          c3
information processing, too many choices and too much free-
dom in browse tend to disorient users and increase difficulty
in information seeking.                                                                       i1        i2   i3        i4   i5    i6   i7    i8        i9   i 10   i 11   i 12    i13   i14    i 15   i 16

   It is well documented in the literature that importance
of dimensions changes contextually. Therefore, we propose                                          Figure 1: Hierarchical categorization.
that the interface for multi-faceted browsing should incor-
porate contextual information by adjusting accessibility of
dimensions according to their importance in varied contexts.
We conduct experiments to investigate whether context-                                  results of search need to be browsed.
sensitive interface will increase browsing effectiveness and                                When browsing at a site, people highly depend on the
the empirical results strongly support the advantages of such                           directories or categories provided by website developers to
interface.                                                                              surf the large amount of content. So far, there are two major
                                                                                        kinds of information categorization: hierarchical and multi-
                                                                                        faceted. Hierarchical categorization employs a predefined
1.     INTRODUCTION                                                                     tree structure to classify objects, and users are required to
   The rapidly increasing online population reveals that peo-                           rigidly follow the structure while traversing down the tree
ple are more and more dependent on the Internet to seek in-                             from the root in order to select appropriate attributes at
formation. Browse and search are the two major strategies                               each level (see Fig. 1). Hierarchical categorization is very
for information seeking. Search relies on recall from users’                            natural when subordinate relationship is clear. Many ob-
memory to express a query. The recall process requires users                            jects, however, do not have such property. For example,
to actively retrieve specific word/phrase, and hence demands                             cars have several attributes (or facets), e.g., make, type,
a high level of cognitive loading and processing [1, 6]. In con-                        year, price, etc., which are virtually independent. Using hi-
trast, browse is based on the recognition process, which only                           erarchical categorization would then unnaturally impose a
needs people to passively confirm if the received information                            rigid ordering on the attributes in order to determine the
is the specific fact or the specific words/phrases they are con-                          levels of the attributes in the hierarchy to classify cars. If
cerned, and hence demands less cognitive efforts. In another                             the subordinate relationship defined by the web developer is
word, browse is more user-friendly than search. Therefore,                              different from a user’s perception, the user may not be able
although search engines have prevailed as the primary In-                               to find the information he needs, or else needs to browse the
ternet application [15], people still heavily rely on browse to                         hierarchy back and forth on the tree ladder.
seek information on the Internet, especially when they do                                  In contrast to hierarchical categorization, multi-faceted
not have much background knowledge of target information                                categorization provides multi-dimensional browse (or multi-
or have imprecise information needs. Even when web users                                faceted browse), as each facet can be viewed as a “conceptual
choose search as the major information seeking strategy, the                            dimension” of the information space, and users can choose
                                                                                        flexibly from any arbitrary dimension(s) to browse. For ex-
                                                                                        ample, consider a set of items i1 , i2 , . . . , i12 with three at-
                                                                                        tributes a, b, and c. Attribute a has terms (values) {a1 , a2 },
                                                                                        b has {b1 , b2 }, and c has {c1 , c2 , c3 } (see Fig. 2). Then, a user
Permission to make digital or hard copies of all or part of this work for               can browse the item set from any dimension he wishes, and
personal or classroom use is granted without fee provided that copies are               in any arbitrary order. For example, he may start with a1 of
not made or distributed for profit or commercial advantage and that copies               attribute a, then b1 of b, and then c1 of c. This browse path
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
                                                                                        (p1 ) leads him to the set {i1 , i2 }. Note that the combinations
permission and/or a fee.                                                                of a1 , b1 , and c1 all lead to the same item set, regardless of
Copyright 200X ACM X-XXXXX-XX-X/XX/XX ...$5.00.                                         their orderings. So if one is more familiar with attribute c,


                                                                                    1
                                  p1             p3                                                facet importance in different contexts. This is particular
                                                                                                   useful when the number of possible facets is large, as taking
                                       a1                                                          context into account may help remove some facets that are
                                                                                    p4             less relevant to contexts.
                                                             b1
                                                                                                      There are two reasons why we would like to investigate
                                                                                        b2
                                                                                                   the role of context played in browse. First of all, context
               p2            c1                                   a2
                                                                                                   is viewed as an important issue in the research of search,
                                            c2                                                     whereas for browse, which is another major form of infor-
                                                                            c3                     mation seeking, the effects of context have not been seri-
                                                                                                   ously and rigorously studied. Secondly, the effect of context-
                                                                                                   sensitive facet arrangement might seem intuitive and self-
                                                                                                   evident. However, as we have mentioned, this approach is
                      i 10        i 11                                                             indirect and non-intrusive as such users might oversee and
                                                   i5             i6                    i14
                                       i2                                    i9                    neglect its effects.
               i 15          i1
                                                 i 13
                                                             i7                          i16          Therefore, the purpose of this paper is to investigate whether
                                                                       i4                          context-sensitive interface does indeed increase browsing ef-
                                            i3                                   i 12
                                                        i8                                         fectiveness for web sites of multi-dimensional categorization.
                                                                                                   A contextualized browsing behavior model is, hereby, estab-
       Figure 2: Multi-faceted categorization.                                                     lished, based on which some guidelines of attribute arrange-
                                                                                                   ment can be provided for interface design on multi-faceted
                                                                                                   categorization systems. We begin by conducting a norm-
                                                                                                   ing experiment to look for the preferences users exhibit in
he may start with c1 to browse his target (e.g., path p2 in the                                    choosing facets under different browsing contexts. We take
figure). Moreover, selecting a partial set of attributes, e.g.,                                     the attributes of cars (from 2000 to 2006 in American car
path p4 , yields a set of items with the specified attributes.                                      market) as our experiment material for these two reasons:
   As illustrated above, multi-faceted categorization treats                                       (1) most people are not automobile experts so they will rely
each attribute equally, and allows users to slice and dice                                         on browse more than search to seek information of cars, and
the information space from any chosen “dimensions”. It                                             (2) automobiles have many attributes that might overwhelm
therefore offers myriad browsing paths and great flexibility                                         people. The browsing interface used in this research is sim-
in the browsing order through different dimensions, as op-                                          ilar to Flamenco but different with the respect of attribute
posed to traditional hierarchical-style of browse where the                                        arrangement.
sequence of attributes selected by users must follow some                                             Our experimental results indicate that browsing behaviors
pre-determined order. Indeed, multi-faceted categorization                                         are indeed context sensitive and that the click frequencies
has already been used and attracted much attention, both                                           of attribute vary according to contexts. On the other hand,
in academia (e.g., Flamenco [23]) and in practice (e.g., En-                                       some attributes remain significantly important than others
deca [10]).                                                                                        regardless of contexts. Moreover, placing attributes more
   However, even though flexibility of multi-faceted catego-                                        accessible to users can further boost click frequencies of at-
rization releases web users from a rigid structure, it, on the                                     tributes which are already important to the context, but
other hand, might offer more freedom than what web users                                            attribute arrangement has little impact on attributes which
can afford. Due to the limit of human information processing                                        are less related to the context. Based on these findings, we
capacity (where the span of our short-term memory is the                                           believe that a good interface design for multi-faceted cat-
magic number seven plus or minus two [14]), a large num-                                           egorization systems should take contexts into account to
ber of attributes might overwhelm users during the process                                         reflect user’s needs and ease his browse efforts from many
of browsing. This phenomenon would interfere information                                           choices. That is, we should place contextually important
seeking and decrease efficiency and effectiveness of browsing.                                        attributes more accessible to users so that they can eas-
So the purpose of this research is to investigate a possible                                       ily and quickly choose the appropriate attributes to filter
solution to this problem of multi-dimensional categorization.                                      out the information they need. Based on these findings we
   We propose that the arrangement of attributes should                                            propose a context-sensitive interface that can arrange at-
be adjusted to user’s information needs. In particular, the                                        tributes dynamically according to the contexts. We then
attributes perceived to be more important should become                                            conduct an experiment to compare this new interface with
more accessible in the interface than attributes of less im-                                       a baseline interface that is “context-insensitive” (whereas
portance. There are many advantages of such adjusted ar-                                           both interfaces are all based on multi-dimensional catego-
rangement of attributes. First of all, it provides guidance                                        rization). Results of the experiment suggest that the new
for users to browse among attributes instead of leaving them                                       interface design allows users to find information more easily
roaming around. Secondly, this kind of guidance is indirect                                        and more effectively, thereby confirming our hypothesis that
and not as intrusive as recommendation [2] or change layout                                        context-sensitive display of attributes can improve browse
constantly. Thirdly, it still preserves the flexibility of multi-                                   effectiveness.
dimensional categorization that allows users to freely switch                                         The rest of the paper is organized as follows. Section 2
to any attribute at any time. To achieve these advantages,                                         presents literature review. Section 3 describes methodology
we need to know user’s needs when browsing information.                                            and results of the norming experiment. Based on the ex-
   As we know, the behavior of browsing occurs in some sort                                        perimental results we present a context-sensitive interface
of context, such as time, place, characteristics of the user,                                      for multi-dimensional browse and conduct an experiment to
needs of the user, etc. The context may in turn affect user’s                                       compare this new interface with a baseline interface that is
ratings of attribute importance; that is, different contexts                                        context-insensitive in Section 4. Finally, Section 5 concludes
yield different attribute importance to users. Thus, it is                                          and provides some direction for future work.
logical to infer that browse can be made more efficient and
effective by arranging the display of the facets according to


                                                                                               2
2.   RELATED WORK                                                     ranged in the window so that users can quickly choose a
   Web search engine used to be characterized as “one size            facet according to his needs, rather than go through the
fits all.” This means the same query formulated by different            entire window to pick up one. To answer this, we need to
users for different purposes would definitely locate the same           know the correlation between facet importance and contexts
search results. However, more and more attentions have                (user’s purchasing needs in this research). Their correlation
been directed to contextual retrieval, in which user prefer-          should be constructed in order to design context-sensitive
ence, needs, or search context are part of the search frame-          interface for multi-dimensional browse. There will, thus,
work. Techniques such as relevance feedback [16, 20] and              be two stages in our research design: norming experiment,
interaction history [12, 22] can provide contextual informa-          and interface comparison experiment. The norming experi-
tion to the system to improve search effectiveness.                    ment is aimed to find out the importance weighting of facets
   In addition to search, browse is another primary way for           across various contexts, which will in turn be used for our
information retrieval. However, to our knowledge, the issue           context-sensitive interface design. The interface compari-
of context has not been discussed for browse in the literature.       son experiment is conducted to compare context-sensitive
Given the essential differences between search and browse,             and context-insensitive interface to see their effects on users’
as reviewed in the previous text, it calls for more empirical         browsing effectiveness. In the following sections we present
research to bear upon this issue.                                     the two experiments.
   Browse relies on categorization as such that people choose
category terms (i.e. dimensions or facets) to narrow down             3. NORMING EXPERIMENT
the information set rather than designating queries. Is it              As we have discussed so far, importance of some facets
true that users’ choosing patterns of categorical dimensions          could be influenced by contexts. Nevertheless, some central
is also influenced by contexts? If it is, then we can in turn          facets could always remain high importance regardless of
use this contextualized browsing model in designing a more            the contexts. For example, consider buying a car. Most
effective interface. Before we proceed to do an empirical              of the time we will think over the manufacturer and price
study for this, the research findings in cognitive psychology          despite of our special preferences. However, if one likes to
may provide some enlightenment.                                       speed, power-train attributes would be more important to
   It is well documented in cognitive psychology that features        him; but if one enjoys luxury, exterior and interior features
or attributes of a category have different degree of impor-            would be preferred. Based on this, we propose the following
tance, which is called graded structure [4, 5]. Most impor-           two propositions:
tantly, the salience of a feature may change with contexts,
such as the receivers or with the external conditions [8].            Proposition 1: Some conceptual dimensions are context
Moreover, the research of Roth and Shobenal [17] also strongly            insensitive. In particular, some conceptual dimensions
suggested that different instances of one particular concept               are always important to a user regardless of his brows-
(category) have different level of salience in different con-               ing context.
text. In another word, the graded structure of categories
is highly dependent on contexts. Based on the above find-              Proposition 2: Some conceptual dimensions are context
ings, we believe that the importance of some attributes of                sensitive; that is, different browsing contexts yield dif-
categories under browsing should similarly alter in different              ferent weights to a conceptual dimension.
contexts. So we can make browse more effective by dynami-
cally increasing accessibility of those important attributes in         We conduct a norming experiment to assess these two
accordance with their contexts, rather than giving all con-           propositions. In this section, we describe methodology and
ceptual dimensions the same weight in all contexts.                   results of the norming experiment. The results can help
   When confronting a large amount of information, peo-               us create a more accurate user model, thereby to design
ple typically want to catch the most important information            a user-centered and context-sensitive browsing interface for
first, and then surf other helpful information if they need to.        multi-faceted categorization systems.
Therefore, in multi-faceted categorization systems, the se-           3.1 Design
quence of facets browsed by users may be influenced by the
importance of facets, which are usually displayed in some                There are two independent variables: display order and
order on the system browse interface (we refer to this part           context, and one dependent variable: attribute importance
of facet listing as the attribute display window, see the left        in the norming experiment. Attribute importance is as-
frame in Fig. 3). A good interface design should, hence, ar-          sessed by frequency of clicks on the attributes. A higher
range the display in a user-centered style so that each user          frequency indicates a higher level of attribute importance.
can efficiently and effectively find the information they need.           However, frequency of clicks can be influenced by order po-
   In particular, we would like to emphasize the “process-            sition of an attribute on the display window. Usually, the
oriented ” nature of the facet arrangement. The arrange-              first couple of attributes in an array of features are more
ment does not require a fixed order of browse, i.e., from the          likely to be clicked than other features positioned later in a
top to the end. On the other hand, it still allows users to ac-       sequence. Therefore, three orders of features are arranged
tively browse through different attributes according to their          such that every feature has equal chance to appear in the
information needs. Most importantly, the arrangement can              top, middle and bottom areas of a sequence of features on
give users an indirect, and personalized guidance by incor-           the display window. The details of order arrangement will be
porating contextual information into the facet arrangement.           illustrated in Section 3.2. The variable of order is a between
Moreover, this kind of process-oriented information seeking           subject variable such that every subject only received one of
is not as intrusive as most recommendation systems have               the three orders when they participated in the norming ex-
posed [2]. Owing to these advantages, we propose that con-            periment. The context variable yields 12 scenarios, each of
textualized facet arrangement should significantly increase            which represents a combination of three contextual factors.
browse satisfaction and effectiveness.                                 The scenarios will be further discussed in Section 3.2 as well.
   A natural question then arises: how facets should be ar-           The context variable is a within subject variable and each
                                                                      subject received all these 12 scenarios in the norming.


                                                                  3
3.2 Experiment Material
   In the norming experiment we use automobile information
from 2000 to 2006 in American car market as our content.
The information was obtained from some Internet automo-
tive marketplaces, and in total 8030 records were collected
into our database. The values of some of the attributes in
the records were more likely to be missing than the others,
and were thus deleted for the norming experiment. In total,
45 car attributes are left to serve as our experimental mate-
rial (see Table 4 in the appendix), which are in turn arranged
into three orders to avoid subjects from using only the first
few attributes to filter the car information. We first cluster
these attributes into 11 groups by their functional similarity
(e.g., front brake, rear brake, and ABS are related to brak-
ing function), and we randomly order the 11 groups in the
attribute display window. We then divide the attribute dis-
play window into three areas: Top, Middle, and Bottom, and
reorganize the areas to yield three different display orders.
In the three display orders, each attribute group only ap-
pears once in one of the three areas. For example, suppose                   Figure 3: The Flamenco search interface.
there are three attribute groups: G1 , G2 , and G3 . Then
the three arrangements of the groups in the attribute dis-
play window are G1 G2 G3 , G2 G3 G1 , and G3 G1 G2 , as shown        Interface
below:
                   Area           Order
                                                                     In the norming experiment, we propose a browsing inter-
                   Top       G1    G2 G3                             face similar to Flamenco to test our propositions, but make
                  Middle     G2    G3 G1                             some changes that will be further described later. Flamenco
                  Bottom     G3    G1 G2                             search interface opens with an overview of the entire col-
                                                                     lection organized in multiple facets.1 Selecting a category
Context Scenario                                                     gathers an initial result set of matching items, and brings
To understand whether subjects prioritize different concep-           the user an interface like Fig. 3 for evaluation and further
tual dimensions and have different browsing patterns in dif-          refinement. The display consists mainly of three parts: the
ferent context scenarios, we give subjects different scenarios        result set, which occupies the right half of the page; the cat-
and ask them to click on attributes in order to choose cars          egory terms which are listed along the left by facet; and the
suitable for each context scenario. To design the scenarios,         selected category terms, which is shown at the top.
we choose three context factors that are commonly used in               Like Flamenco, our interface first opens with facets along
car market surveys: number of family members, ratio of rev-          with their top-level categories to give users an overview of
enue to expenditure, and purchasing preference. The levels           the entire collection. Users can select a category to gather
of each contextual factor are described in the following:            an initial result set of matching items and enter an interface
                                                                     like Fig. 4, which comprises the category terms on the left,
   • Number of family members: single, extended family
                                                                     the result set on the right, and the selected category terms
   • Ratio of revenue to expenditure: low, high                      on the top. Users spend most of their time in this display on
                                                                     evaluating the results. They can select a category to narrow
   • Purchasing preference: luxury, speed, security                  the results or remove a selected category term to expand
  There is another level “economy” that has been commonly            them. Selecting an individual item will cause all detailed
used to represent purchasing preference. However, we have            information of the item to be displayed.
removed it here because of its correlation with low ratio of            Flamenco search interface is designed for image brows-
revenue to expenditure. In total, the three variables yield          ing. In image browsing, people usually browse what they
12 combinations (2 × 2 × 3), each representing a purchasing          are interested in and do not need to make much compari-
scenario with different needs specified by the three contex-           son between different images. In contrast, car information
tual factors. The scenario is further presented in different          browsing involves decision making. For such kind of brows-
tones, one for female and the other for male. The scenarios          ing, people need to compare between different items to help
are assigned to subjects according to their gender so that           making their purchase decision. For this purpose, we change
subjects of different sex could easily get involved in the sce-       the display of attribute values (category terms) in the at-
nario. The following illustrates one of the purchasing needs         tribute display window. When subjects choose a value of
stated from male’s point of view. The one stated in female’s         one attribute, the other values of that attribute will also
tone is almost the same, except that it starts with “You are         be displayed, unless there is no car meeting the constraint.
a single woman.”                                                     As such, it is easy for subjects to compare between differ-
  Purchasing need—single, high ratio of revenue to expen-            ent choices within one attribute in order to find the most
diture, and purchasing preference to luxury:                         fitting items. For example, in Fig. 4 when subjects choose
                                                                     the max price of “< 50, 000”, this choice is marked by ‘ ’
     You are a single man without any mortgage pressure.
                                                                     and changed to red color. The other possible values of “max
     The basic expenditure is only 1/3 of your income.
                                                                     price” are still over there so that subjects are aware of other
     Recently, you are promoted to be general manager,
                                                                     possible choices all the time during their browse. They can
     so you wish to buy a car fitting in with your status.
                                                                     quickly switch to other choices if they want to, therefore to
     You also wish to have a luxury car to attract people’s
     attention. (...man’s point of view)                             1
                                                                         See http://bailando.sims.berkeley.edu/flamenco.html.


                                                                 4
                                                                    to browse the information shown on the screen and pick out
                                                                    some cars they would consider to buy. They can browse
                                                                    as much as they want and can stop when they feel like to
                                                                    proceed to the following scenario. The program records the
                                                                    frequency of clicks on each of the 45 attributes. The results
                                                                    are hereafter used to design the context-sensitive interface.

                                                                    3.5 Results of Norming Experiment
                                                                    3.5.1 Experimental model
                                                                       Due to the fact that there are both between-subject and
                                                                    within-subject effects in our norming experiment, we choose
                                                                    to use the split-plot factorial design as our experimental
                                                                    model [13]. We refer to the design as SPF-p · qrst. The
                                                                    parameter p represents the levels of display ordering, which
                                                                    is the only between-subject effect in the model, and it has
                                                                    value p = 3. The three parameters q, r, and s represent
                                                                    the levels of the three context variables number of family
                                                                    members, ratio of revenue to expenditure, and purchasing
                                                                    preference, respectively. Their values are q = 2, r = 2,
                                                                    and s = 3. All the three variables are within-subject. Fi-
                                                                    nally, the parameter t represents the levels of attribute, also
                                                                    within-subject because every subject receives 45 attributes
                                                                    in the norming experiment. The value of t is set to 29. One
Figure 4: Main browsing interface of this research.                 may recall that there were originally 45 attributes. Sixteen
                                                                    of them were removed because they were hardly clicked by
                                                                    subjects, meaning that these attributes were not considered
easily make comparison between different choices.                    as relevant in any of the 13 scenarios. Thus, our experimen-
                                                                    tal model is SPF-3·2 2 3 29 with each subject as a block.
3.3 Subjects                                                        3.5.2 Analysis of variance
  162 subjects (81 male and 81 female) participated in the
experiment as paid volunteers. They were recruited from                The results of the norming experiment are shown in Ta-
advertisements posted on BBS in various universities of the         ble 1.2 We have two important findings in the results. First,
metropolitan Taipei area. The average age of those subjects         the variable of attribute has significant main effect on the
were 22.2. We also asked them some questions regarding              frequency of clicks (p < .001). That is, some attributes are
their experiences about cars. Out of 162 subjects, 85 of            always clicked more often than others across all the different
them (52%) could drive, and among those who could drive,            scenarios. The finding strongly supports our first proposi-
the average number of drive year was 3.6. Moreover, we              tion which states that some conceptual dimensions are al-
asked them how familiar they were with cars. Sixty-four             ways more salient under different browsing contexts.
of them (39%) rated themselves as novice, and 89 (55%)                 The Scheffe’s post-hoc test on the main effect of attribute
rated themselves having some knowledge about cars, and 9            is conducted, and we find that click frequencies of the follow-
(5%) persons rated themselves as expert. In addition, 113 of        ing attributes, Type, Seat Capacity, Make, and Max Price
them (70%) had at least one car in their family. Even though        are significantly higher than click frequencies of other at-
the subjects were college students, about half of them could        tributes (p < .001). That is, Type, Seat Capacity, Make, and
drive and about 70% of them had at least some knowledge             Max Price appear to be the most important considerations
about cars. Therefore, this sample was representative to            when people want to purchase a car in different contexts.
some extent.                                                        This helps us to understand that these four attributes are
                                                                    important regardless of contexts, and they should be given
3.4 Procedure                                                       a high priority in ordering across different contexts.
   Subjects are randomly assigned to one of the three dis-          Effects of Contexts
play orders, with 54 subjects in each, balanced in male and
female. Every subject participates in the experiment on an          The second major finding of the norming experiment is about
individual basis. Moreover, recall that the three context           the effects of context. The results show that attribute and
variables produce 12 different scenarios. Each subject in            each context variable have significant interaction effect on
the norming experiment is given 4 of the 12 context sce-            2
                                                                       In ANOVA, the F statistics is used to test a hypothe-
narios plus one training task. The scenario of the training         sis about the effect. It is the ratio of two mean squares,
task is designed to be very different from any of the 12 con-        MSA/MSE. MSA is for the effect and MSE is for the error.
text scenarios so that the learning effects of the system can        Because the mean squares in our model do not have ap-
be avoided. Subjects are asked to read the description of           propriate expected values for constructing F statistics, we
                                                                    use the composite mean squares to construct the quasi-F
the experiment on the screen and then proceed to read the
                                                                    statistics [13], denoted by the symbol F . The sampling dis-
context scenario description. The task of the subjects is to
browse automobile information provided by our system so             tribution of F is the F distribution with v1 and v2 degrees
                                                                    of freedom for MSA and MSE, respectively. The F distribu-
as to choose suitable cars for each context scenario. Sub-          tion can be used to determine the probability p of observing
jects are allowed to take their pace in doing all these tasks       an F statistic larger than or equal to that obtained. If p is
on a computer program which displays scenarios and the              less than 0.05, it is considered to have sufficient evidence for
browsing interface of the car database. They are required           significant effect.


                                                                5
                                                                      tributes and efficiently screen out the information he does
                Table 1: ANOVA results                                not need.
 Effect                    F (v1 , v2 )            MSE       p
                                                                         In this section we incorporate these design guidelines and
 Attribute                F (28, 95.46) = 5.380   7.560   0.000
 Attribute×Number of      F (28, 11.36) = 2.612   0.834   0.005
                                                                      propose a context-sensitive interface for our automobile in-
 family members                                                       formation system. We then conduct an experiment to com-
 Attribute×Ratio of       F (28, 11.36) = 5.523   0.399   0.001       pare this new interface with the baseline interface which
 revenue to expenditure                                               simply arranges attributes in an arbitrary order.
 Attribute×               F (56, 26.83) = 7.945   0.398   0.000
 Purchasing preference                                                4.1 Context-Sensitive Interface Design
 Order                    F (2, 12.99) = 0.123     2.82    0.89          Our new interface is similar to the one used in the norm-
 Attribute×Order          F (56, 24.72) = 3.161   0.473   0.000       ing experiment, except in the layout of the attribute display
                                                                      window, where we make two modifications. The first one is
                                                                      to adjust the attribute ordering. Recall from Section 3.5 that
the click frequencies of attributes (p < .001, except “At-            we have identified attributes that are always important re-
tribute × Number of family members” is significant at p <              gardless of contexts, and attributes whose importance varies
.01). These interaction effects support our second proposi-            according to contexts. We arrange important attributes that
tion which asserts that attribute’s weight vary according to          are “context-independent” on top of the attribute display
contexts. Next, we examine which attributes have weights              window, followed by the attributes whose importance are
dependent on contexts. The information will help us distin-           contingent upon contexts.
guish attributes in designing our context-sensitive interface.           The second modification is to reduce the size of the at-
   The results of Scheffe’s post-hoc test on the interaction           tribute display window, so as to alleviate user’s efforts in
effects of attribute and context variables show that contexts          locating the attributes they need (recall that there are 45-
determine the importance of attributes. For example, we               attributes plus their respective values!). To do so, the at-
find that subjects pay more attention to seat capacity and             tributes are displayed in two modes: “expanding” and “col-
safety related attributes when the scenario is in an extended         lapsing”. The expanding mode shows all possible values of
family (p < .001). Moreover, subjects focus on Type, Secu-            the attributes, whereas the collapsing mode only shows the
rity System, Driver Front Airbag, Make, and Price in sce-             name of the attributes and users need to expand them to
narios of high ratio of revenue to expenditure (All results are       see the respective values. For the important attributes (e.g.,
statistically significant at p < .001, except Security System          make, max price, type, seat capacity, and engine) whose im-
and Driver Front Airbag are significant at p < .05). Fi-               portance is either context free or context dependent, they
nally, the experiment results indicate that purchasing pref-          are put in the expanding mode by default such that users
erence to luxury causes high click frequency of interior fea-         can see the attributes and their values all at once. Other
tures (p < .001); purchasing preference to speed causes high          attributes (e.g., year), are put in the collapsing mode, unless
click frequency of tire related attributes and engine related         the user explicitly expands them.
attributes (p < .001); while purchasing preference to secu-              We refer to the interface with dynamic attribute order-
rity causes high click frequency of safety and security related       ing as Contextually-Ordered Attribute Display (COAD). For
features (p < .001). The results strongly supports Proposi-           comparison, we use a baseline interface that simply dis-
tion 2.                                                               plays attributes in some arbitrarily fixed order, referred to as
                                                                      Randomly-Ordered Attribute Display (ROAD). In the inter-
Effects of Orders                                                     face comparison experiment, subjects are randomly assigned
We also have other interesting findings about orders. Even             to one of these two interface designs (COAD vs. ROAD).
though the result (row 5) in Table 1 shows that ordering does         Four dependent variables will be assessed to see the effects
not have a clear impact on attribute’s overall click frequency,       of the two different interfaces: ease to find information, con-
attribute and its ordering have significant interaction effect          fidence about their decision making quality, ease to browse,
on an attribute’s click frequency (p < .001). By examining            and overall satisfaction.
the experiment data, we found that the attributes that are
affected by their orderings are those that are already impor-          4.2 Hypotheses of Interface Comparison Ex-
tant to the context. Yet, attributes that are less related to             periment
the context remain lower click frequencies even if displayed            Since conceptual dimensions in COAD are arranged based
on the top of the attribute display window. This result fur-          on user’s purchasing needs in contexts, we believe that such
ther sheds exciting light on the essential role of contexts.          a context-sensitive interface will allow the users to browse
                                                                      more effectively and with more satisfaction. So the following
                                                                      four hypotheses are proposed and the interface comparison
4.   INTERFACE COMPARISON EXPERIMENT                                  experiment is conducted to test these hypotheses:
   The results of norming experiment confirm our two propo-
sitions on user’s browsing behaviors; that is, some attributes        Hypothesis 1: The average rating of ease to find informa-
are significantly more important than others regardless of                tion for subjects in the COAD interface condition will
contexts, while some may vary from context to context.                   be higher than that of subjects in the ROAD interface
Moreover, we also learned that the ordering of attributes                condition.
in the attribute display window can boost the click frequen-          Hypothesis 2: The confidence level of decision making qual-
cies of attributes which are important to the context, but               ity of subjects in the COAD interface condition will
has little impact on the attributes which are less related to            be higher than that of subjects in the ROAD interface
the context. Based on these findings, we believe that a good              condition.
interface design in multi-dimensional browse should reflect
these contextual effects. That is, we should place contex-             Hypothesis 3: The average rating of ease to browse for
tually important attributes more “accessible” to the user                subjects in the COAD interface condition will be higher
so that he can easily and quickly choose the important at-               than that of subjects in the ROAD interface condition.


                                                                  6
Hypothesis 4: The level of overall satisfaction of subjects          4.5 Subjects
   in the COAD interface condition will be higher than                  73 subjects (43 male and 30 female) were recruited in the
   that of subjects in the ROAD interface condition.                 same way as norming experiment to participate in this test
   In another word, it is hypothesized that subjects are more        as paid volunteers. The average age of those subjects were
satisfied, feel easier to browse and find needed informa-              21.8. We also asked them some questions regarding their ex-
tion, and feel more confident about their decision making             periences about cars. Out of 73 subjects, 32 of them (48%)
quality when using context-sensitive interface (COAD) to             could drive, and among those who could drive, the average
browse than when they browse in context-insensitive inter-           number of drive year was 2.6. Moreover, we asked them how
face (ROAD).                                                         familiar they were with cars. Twenty-three of them (32%)
                                                                     rated themselves as novice, and 49 (67%) rated themselves
4.3 Design of Interface Comparison Experi-                           having some knowledge about cars, and 1 (1%) person rated
    ment                                                             himself as expert. In addition, 57 of them (74%) had at least
                                                                     one car in their family. Even though the subjects were col-
   The study employs two independent variables and four de-          lege students, almost half of them could drive and about 68%
pendent variables. The independent variables are interface           of them had at least some knowledge about cars. The de-
(context-sensitive interface (COAD) vs. context-insensitive          mographic information collected for the interface experiment
interface (ROAD)) and context. COAD interface is used                was similar to those collected for the norming experiment.
for subjects in the experimental group, while the baseline           Therefore, the sampling method should be appropriate and
interface ROAD is used for the controlled group. The inter-          the sample should be representative.
face variable is a between-subject variable and each subject
is randomly assigned to only one of the two conditions of            4.6 Results of Interface Comparison Experi-
interface. The context variable is a within-subject variable.            ment
Each subject is given 6 context scenarios plus one training
task to complete. The context scenarios will be further dis-            We analyze the data collected from the experiment in
cussed in Section 4.4.                                               two parts. One is for the two questions answered at the
   The experiment procedure is the same as the norming ex-           end of each scenario and the other is for the satisfaction
periment, except that there are two questions at the end             questionnaire. For each question of the questionnaire, if
of each scenario, one concerning ease of finding needed in-           the responded value is above or below two standard devia-
formation, and the other concerning confidence on selected            tions away from the mean, it will be viewed as the outlier
items. Moreover, after finishing all tasks, subjects are asked        and be deleted from the analysis. For the questions after
to fill out another questionnaire to evaluate their rating of         each scenario, 26 out of 876 responses (73*6*2) are outliers,
ease to browse and overall satisfaction.                             which only occupies 2.97% of the total responses. As for
                                                                     the two questions of Ease to browse, 13 out of 146 (73*2)
4.4 Experiment Material                                              responses are outliers (7 for COAD and 6 for ROAD); and
                                                                     for the three questions of overall satisfaction, 13 out of 219
Context Scenario                                                     (73*3) responses (6 for COAD and 7 for ROAD) are out-
                                                                     liers. Roughly equal numbers of outliers are deleted from
In this experiment we still use the three context variables to       both interface conditions, such that the deletion of outliers
design the context scenarios. However, from the norming ex-          should not affect the results.
periment results, we find high correlation between “extended
family” of Number of family members and “security” of pur-           4.6.1 Results of Ease to Find Information and Deci-
chasing preference. Therefore, we remove security from the                 sion Making Confidence
purchasing preference and the levels of the three context
variables are as follows:                                               We again use the split-plot factorial design as our exper-
                                                                     imental model because there are both between-subject (in-
   • Number of family members: single, extended family               terface) and within-subject effects (context scenario) in this
                                                                     experiment. We refer to the design as SPF-p · qr. The three
   • Ratio of revenue to expenditure: low, high                      parameters p, q, and r are defined as follows. The param-
   • Purchasing preference: luxury, speed                            eter p = 2 represents the two different interfaces (COAD
                                                                     vs. ROAD), which is the only between-subject effect in the
  The three variables allow us to construct 8 different sce-          model. The other two parameters q = 6 and r = 2 are for
narios, among which, we have further removed two scenarios           two within-subject effects: numbers of contexts and num-
concerning “extended family” in combination with “speed”             bers of questions. Thus, the experimental model is SPF-2·6
purchasing preference. This is because many subjects re-             2 with each subject as a block.
sponded that these two scenarios are not reasonable and so              Analogous to how we analyze in the norming experiment,
they were not sure what kind of cars were suitable for the           we use the ANOVA procedure to analyze the effects in our
scenarios. Hence, only 6 scenarios were used in the test.            experimental model. In the ANOVA procedure, attribute
                                                                     display, context, question, and subject, as well as their in-
Satisfaction Questionnaire                                           teractions, are independent variables. We perform the Stu-
The post-test satisfaction questionnaire is designed to mea-         dent’s t-test on the interaction effect of interface and ques-
sure subject’s overall satisfaction with the interface they          tion to see if there are significant differences between these
receive. Questions in the questionnaire measure two con-             two interfaces for both of the two rating questions.3
cepts, which are later constructed through exploratory fac-          3
tor analysis. The first concept is subject’s perception of              The t statistics is used to test a priori hypothesis about the
ease of browse in the system, and the second concept is sub-         mean difference between two groups. It is the ratio of mean
                                                                     difference to the square-root of the estimated variance of
ject’s satisfaction with the arrangement of attributes. All                                           σ
                                                                     mean difference, (m1 − m2 )/ˆij . The sampling distribution
the questions in this experiment employ a measure of the             of t is the t distribution with v degrees of freedom for MSE.
7-point Likert scale.                                                The t distribution can be used to determine the probability p


                                                                 7
                                                                                 icantly important than others regardless of contexts. More-
Table 2: Student’s t-test on the interaction effect of                            over, placing attributes more accessible to users can boost
interface and question.                                                          an attribute which is already important to the context, but
 Comparison                     Mean             t(v)       ˆ
                                                            σij       p
                              Difference                                          has little impact on an attribute which is less related to
 (Rating on Q1 in COAD)        0.3437       t(36.89)       0.1047    0.001       the context. Based on these findings we then proposed a
 −   (Rating on Q1 in ROAD)                 = 3.281                              context-sensitive interface prototype that can dynamically
 (Rating on Q2 in COAD)         0.2272      t(36.89)       0.105     0.019       place important attributes at a more “accessible” position
 − (Rating on Q2 in ROAD)                   = 2.164
                                                                                 for users to choose. We then conducted another experiment
                                                                                 with a baseline interface that is context-insensitive. The
                                                                                 results showed that the new interface design allowed users
Table 3: ANOVA results for ease to browse and                                    to find information more easily and more effectively, and
overall satisfaction                                                             with more satisfaction to their needs. The empirical re-
      Effect                   F (v1 , v2 )         MSE         p
                                                                                 sults converge to confirm our hypotheses. In particular, the
      Ease to browse          F (1, 71) = 9.81     0.297    0.0025
      Overall satisfaction    F (1, 71) = 1.37      1.17     0.25                context-sensitive interface used in this study can indeed im-
                                                                                 prove browse effectiveness, even though it only provides indi-
                                                                                 rect and non-intrusive arrangement of facets. These results
                                                                                 suggest legitimacy of further rigorous studies on the effects
   Table 2 shows that both questions Q1 (ease to find in-
                                                                                 of context for mult-dimensional browse, and more investi-
formation) and Q2 (confidence in decision making quality)
                                                                                 gations on possible mechanisms to incorporate contextual
are rated significantly higher in the COAD interface than in
                                                                                 information into interface design.
the ROAD interface. That is, subjects in context-sensitive
                                                                                    Our research is closely related to recommendations in on-
interface condition (COAD) feel significantly easier to find
                                                                                 line shops (e.g., [18, 3, 7]), but distinguished by the follow-
information they need than those subjects in the context-
                                                                                 ing important characteristic: their recommendation schemes
insensitive interface condition (ROAD; 5.37 vs. 5.03)). More-
                                                                                 are result-oriented, while ours is process-oriented. That is,
over, the results also demonstrate that subjects feel more
                                                                                 rather than recommending the results directly to users, we
confident with their decision making of cars when they are
                                                                                 choose to let users “participate” in the process of getting the
in the context-sensitive interface condition than those in the
                                                                                 final results so that they can be more confident on and even
context-insensitive interface condition (5.32 vs. 5.10). The
                                                                                 learn more from the information they acquire. At the same
empirical results strongly support hypotheses 1 and 2.
                                                                                 time, by using this context sensitive interface, we provide
4.6.2 Overall Satisfaction                                                       non-intrusive, indirect guidance in the process, which signifi-
                                                                                 cantly facilitates users to browse more effectively. Moreover,
   In the post-test satisfaction questionnaire we use the ANOVA                  this process-oriented interface also provides users the oppor-
procedure to analyze if these two different interface designs                     tunity to explore new world and discover serendipity during
have significant effect on the concepts of ease of browse and                      the process of browsing, which a result-oriented approach
overall satisfaction. In the ANOVA procedure, interface is                       could not offer.
the independent variable, and the ratings of the two con-                           The results of this study have shown the importance of
cepts are dependent variables. Table 3 shows the results.                        context. However, the three contextual factors (i.e., num-
   In Table 3, we can see that interface has significant main                     ber of family members, ratio of revenue to expenditure and
effect on ease to browse (5.45 vs. 5.05; p = 0.001). Subjects                     purchasing preference) used in this study are not the only
do think it is easier to browse when they are in the context-                    contextual factors that are critical for car purchasing deci-
sensitive (COAD) interface. As for overall satisfaction, the                     sions, or even for other kinds of decisions. We need to do
effect of interface is not significant (5.02 for COAD vs. 4.72                     more research to assess the importance of various contex-
for ROAD; p = 0.251). This result might be due to slow re-                       tual factors for browse. Moreover, we also need to find out
sponse time of our system, which ranges 4-7 seconds for each                     how to solicit or extract contextual information, either di-
click. The slow response time might override positive effects                     rectly from users or indirectly from database. Research in
of COAD with respect of overall satisfaction. The above re-                      recommendations has explored several techniques, including
sults, nonetheless, show that COAD is consistently superior                      web mining [19, 11] and machine learning [9, 21] to ana-
to ROAD for browsing, and, hypotheses 1, 2, 3 are also sup-                      lyze user preferences. The techniques can help us to ex-
ported by our results. In another word, the results in the                       tract contextual information and can thus be incorporated
interface comparison experiment consistently and strongly                        to make a contextualized browsing interface more sophisti-
demonstrate the effects of context-sensitive interface design                     cated and personalized in “recommending” attributes and
on browsing effectiveness and satisfaction.                                       browse paths to users. This is indeed our ultimate goal—
                                                                                 constructing a contextually multi-faceted metadata brows-
5.    CONCLUSIONS AND FUTURE WORK                                                ing system to assist browse and decision making for online
   In this paper, we proposed a browsing interface which                         shops.
takes contexts into account to assist browse in multi-faceted                       Finally, our research here investigated a simple prototype
categorization systems, and conducted experiments to inves-                      which arranged important attributes on top of the attribute
tigate whether this context-sensitive interface will increase                    display window to provide high accessibility. Other possible
browsing effectiveness. The results strongly support our hy-                      arrangements and displays to “catch” a user’s attention, e.g.,
potheses. First of all, the results of the norming experi-                       graphical visualization, can be explored to improve browsing
ment indicate that browsing behaviors are indeed context                         efficiency and effectiveness.
sensitive, and attribute importance may vary according to
contexts. On the other hand, some attributes remain signif-                      6. REFERENCES
of observing a t statistic equal to or as large as that obtained.                 [1] E. M. Achilles. Experimental Studies in Recall and
If p is less than 0.05, it is considered that there is significant                     Recognition. Johnson Associates, New York, NY,
difference between the two means.                                                      United States, 1920.


                                                                             8
 [2] G. Adomavicius and A. Tuzhilin. Personalization
     technologies: a process-oriented perspective.                          Table 4: Car Information
                                                                           Groups          Car attributes
     Communications of the ACM, 48(10):83–90, 2005.                   Number of people     Type
 [3] A. Ansari, S. Essegaier, and R. Kohli. Internet                                       Seat capacity
     recommendation systems. Journal of Marketing                        Price related     Make
                                                                                           Max price
     Research, 37(3):363–375, 2000.                                                        Min price
 [4] L. W. Barsalou. Ad hoc categories. Memory and                                         Year
     Cognition, 11(3):211–227, 1983.                                     Gear related      Gear type
                                                                                           # of gears
 [5] L. W. Barsalou. Ideals, central tendency, and                         Security        Security
     frequency of instantiation as determinants of graded                                  system
     structure in categories. Journal of Experimental                                      Driver front airbag
                                                                                           Driver side airbag
     Psychology: Learning, Memory, and Cognition,                                          Passenger front airbag
     11(4):629–654, October 1985.                                                          Passenger side airbag
 [6] L. W. Barsalou. Cognitive Psychology: An Overview                                     Rear child safety locks
                                                                            Safety         ABS
     for Cognitive Scientists. Lawrence Erlbaum                                            Rear brake
     Associates, Hillsdale, NJ, United States, 1992.                                       Front brake
 [7] S. M. Bohte, E. Gerding, and H. L. Poutr´.  e                       Tire related      Tire speed rate
                                                                                           Wheel drive
     Market-based recommendation: Agents that compete                                      Type of spare tire
     for consumer attention. ACM Transactions on                                           Tire aspect ratio
     Internet Technology, 4(4):420–448, 2004.                                              Turning radius
                                                                        Engine related     Horse power
 [8] J. S. Bruner, J. J. Goodnow, and G. A. Austin. A                                      Engine liters
     study of Thinking. Wiley, New York, NY, United                                        Cylinder configuration
     States, 1956.                                                                         Torque
                                                                                           Valvetrain
 [9] K.-W. Cheung and L. F. Tian. Learning user                                            # of valves
     similarity and rating style for collaborative                  Entertainment features CD player
     recommendation. Information Retrieval,                                                LCD monitor
                                                                                           Navigation system
     7(3-4):395–410, 2004.                                                                 # of speakers
[10] Endeca. http://www.endeca.com.                                     Seat and trim      Sunroof
[11] H. Ishikawa, M. Ohta, S. Yokoyama, J. Nakayama,                                       Seat upholstery
                                                                                           Cell phone pre-wiring
     and K. Katayama. Web usage mining approaches to                     Fuel related      Fuel economy city
     page recommendation and restructuring. International                                  Fuel economy highway
     Journal of Intelligent Systems in Accounting, Finance                                 fuel tank capacity
                                                                                           Compression ratio
     and Management, 11(3):137–149, 2002.                              Standard specs      Ground clearance (min)
[12] T. Joachims. Optimizing search engines using                                          Cargo volume(max)
     clickthrough data. In KDD ’02: Proceedings of the                                     Exterior length
                                                                                           Exterior width
     eighth ACM SIGKDD international conference on                                         Exterior height
     Knowledge discovery and data mining, pages 133–142,                                   Curb weight
     New York, NY, USA, 2002. ACM Press.
[13] R. E. Kirk. Experimental Design: Procedures for the
     Behavioral Sciences. Brooks/Cole, Pacific Grove, CA,
     United States, third edition, 1995.                        information retrieval. In SIGIR ’05: Proceedings of the
[14] G. A. Miller. The magical number seven, plus or            28th annual international ACM SIGIR conference on
     minus two: Some limits on our capacity for processing      Research and development in information retrieval,
     information. Psycological Review, 63:81–97, 1956.          pages 59–66, New York, NY, USA, 2005. ACM Press.
[15] L. Rainie and J. Shermak. Reports: Online activities  [21] G. I. Webb, M. J. Pazzani, and D. Billsus. Machine
     & pursuits. Retrieved Arpil 2006 from                      learning for user modeling. User Modeling and
     http://www.pewinternet.org/PPF/r/167/report_display.asp , User-Adapted Interaction, 11(1-2):19–30, 2001.
     Nov. 2005.                                            [22] R. W. White, J. M. Jose, C. J. van Rijsbergen, and
[16] J. Rocchio. The Smart Retrieval System-Experiments         I. Ruthven. A simulated study of implicit feedback
     in Automatic Document Processing, chapter Relevance        models. In Proceedings of the 26th European
     feedback information retrieval, pages 313–323.             Conference on Information Retrieval (ECIR), volume
     Prentice-Hall, Englewood Cliffs, NJ, United States,         2997 of Lecture Notes in Computer Science, pages
     1971.                                                      311–326. Springer-Verlag, 2004.
[17] E. M. Roth and E. J. Shoben. The effect of context on  [23] K.-P. Yee, K. Swearingen, K. Li, and M. Hearst.
     the structure of categories. Cognitive Psychology,         Searching and organizing: Faceted metadata for image
     15(3):346–378, July 1983.                                  search and browsing. In CHI ’03: Proceedings of the
[18] J. B. Schafer, J. A. Konstan, and J. Ried. E-commerce      SIGCHI conference on Human factors in computing
     recommendation applications. Data Mining and               systems, pages 401–408, New York, NY, United
     Knowledge Discovery, 5(1-2):115–153, 2001.                 States, April 2003. ACM Press.
[19] C. Shahabi and F. Banaei-Kashani. Efficient and
     anonymous web-usage mining for web personalization.   Appendix
     INFORMS Journal on Computing, 15(2):123–148,          The appendix gives car information used in the experiment;
     2003.                                                 see Table 4
[20] X. Shen and C. Zhai. Active feedback in ad hoc



                                                          9

				
hkksew3563rd hkksew3563rd http://
About