Introduction
Providing users with interactive control over the
recommendation process has only recently started to receive more
attention in Recommender Systems (RS) research [
25, 26,
40
]. In terms of objective error metrics, recommender
algorithms are already quite mature and only small
improvements can be expected from further optimizing algorithmic
precision [
40
]. However, high accuracy is not the only factor
determining user satisfaction [
26
]. It is increasingly
recognized that user-related aspects such as control, trust and
transparency influence the users’ perception of the
recommendations even more, and may contribute considerably
to higher satisfaction [
26, 40
]. This makes it an important
research goal to let users influence the recommendation
process and to make it more comprehensible [
25, 26, 40
].
Several models exist that describe typical user behavior
during the recommendation process. In earlier work [
31
], for
instance, we have proposed a model comprising three
interaction loops, which represent a) the user’s interaction with
the recommendations themselves, b) selection and
weighting of properties related to the recommended items, and
c) adaptation of entire recommender applications. Various
models have also been introduced in the area of
information retrieval, particularly aiming at examining the users’
information-seeking behavior [
28, 34
]. Due to their focus
on document collections and explicit search tasks, these
models are however not directly applicable to RS. On the
other hand, models in the area of RS research often focus
on conversational and critique-based systems [
44, 8
], more
basic feed-back processes [
41
], or describe system usage
distinguished by different feedback types [
22
], i.e. ways to
elicit implicit or explicit rating data. In [
9
], the area of
interactive RS is surveyed by means of a basic model
comprising those recommender components that can be extended
to allow for additional interaction. While similar in some
aspects to the framework we propose in this paper, the focus
of the authors lies on visualizations and related aspects.
How to offer users more control at the different stages in
the recommendation process is only one of many aspects
mentioned.
In this paper, we will therefore provide a closer look at this
particular issue: First, we present a framework of
interaction in RS that describes the range of possibilities users
have for influencing the recommendation process. Next,
we provide a detailed overview of the four aspects we have
identified around the recommendation algorithm itself that
allow for integrating additional interaction-input data, user
model, external context model and presentation. We survey
some of the most influential work related to each aspect,
derive future research challenges, and outline solutions to
deal with them that are especially promising from our point
of view and subject of our upcoming work. Finally, we
conclude the paper with a short summary and discussion.
A framework for interactive recommending
Figure 1 shows our proposed framework: Blue boxes
represent components containing data, models, or presentation
that may be manipulated by the user to adapt the system’s
outcome according to his or her current needs. The
central recommender algorithm(s) (red circle) that process
input data and models may also be interactively influenced,
for ex-ample, by changing an algorithm’s parameters or by
rear-ranging the processing steps in the case of hybrid
systems.
All of these components can be considered important with
regard to user-perceived quality [
25, 26, 40
], e.g. perceived
recommendation quality or transparency of the results.
There have indeed been efforts to allow users to
manipulate the recommender algorithms themselves [
20
], to
choose from different algorithms [
14
], or to change their
influence in hybrid settings [
6, 30
]. However, in the following,
we concentrate on a) input data related to users or items
provided for the recommender, b) the user model inferred
from, e.g., the user’s preferences, needs, and emotions, c)
the external context model representing the user’s current
situation, i.e. his or her environment, used device, etc., as
well as d) the presentation of the recommender’s results.
For each aspect, a (nonexhaustive) list of properties is
presented which may characterize the respective part of the
system. Arrows (orange) visualize the process flow starting
from possible preprocessing steps and selection of
appropriate input data for the algorithms, which then generate
the recommendations, i.e. adapt the presented result set.
Therefore, the algorithms are able to exploit user model and
external context model, which in turn may be inferred by
means of the users’ feedback or are generally affected by
their interaction with the system.
Current position and future work
Although much effort has been put into improving the
algorithms used in RS, other aspects still lack attention from
the research community, especially regarding their role in
increasing the recommenders’ transparency and the users’
influence on the systems [
25, 26, 40
]. In the following, we
therefore have a closer look at the four relevant aspects
from our model, related work and future challenges.
Input Data
The input for a RS, i.e. user or item data, is not only used
by machine learning techniques to generate
recommendations, but also represents an important part of such a
system that might be exploited to let users influence the
recommendation process and to improve their
understanding of why certain items are recommended.
Collaborative Filtering (CF), the most frequently used
recommender technique [
42
], relies on input data usually
limited to user feedback, which is either explicitly provided
through ratings or implicitly observed based on behavioral
data [
22
]. Other methods use tags [
43
] or rely on a social
graph, i.e. relationships between users [
17, 18
].
Particularly in content-based filtering [
11
], item attributes or other
content- related information are used to recommend items.
However, in all cases, user or item data primarily serve as
input for the algorithms that generate recommendations.
Only few methods exploit, for instance, tags [
12, 46
] or item
attributes [
30
] to let users select and weight certain
product characteristics, or visualize social connections [
17
] to
improve users’ understanding of the recommendation
process.
Eliciting user preferences is an important step in order
to obtain the input data necessary for the employed
algorithms, which is especially relevant in cold-start
situations. Various methods have been proposed to overcome
the problems of traditional rating-based interfaces. Prior
research has shown that ratings may be inaccurate [
2
]
and that users prefer com-paring items instead of rating
them [
23
]. In general, different users seem to benefit from
different interaction possibilities [
24
]. Thus, we among
others have proposed alternative preference elicitation
methods: Our choice-based approach [
32
] allows users to
state their initial preferences without the need to rate items.
When compared to a conventional rating pro-cess, it has
been shown to be more beneficial in terms of, e.g.,
perceived effort, control, and subjective recommendation
quality [
32
]. Other authors have also experimented with novel
ways to elicit preferences, for example, by letting users pick
from groups of items [
7
] or by mapping their choice of
certain pictures to factors describing their preferences [
37
].
We argue that exploiting input data for purposes other
than feeding them into the algorithms can be an important
means for giving users more control over the
recommendation pro-cess. A possible challenge for future research
can therefore be seen in developing techniques that create
new ways of interacting with user or item data. This may
comprise filtering these data even before applying the
algorithms or visualizing them in order to improve the user’s
understanding of product space and his or her position inside
it (as it has been done, for instance, through maps
showing a “recommendation landscape” [
16
]). By building on the
aforementioned works, we particularly want to improve
preference elicitation for CF: Providing alternatives to simply
rating a set of items seems to be a promising way to
alleviate the cold-start problem [
32, 37, 7, 13
]. Now imagine
an extension of [
32
] that provides users with comparisons
that not directly feature the items (presented in form of, e.g.,
movie posters, hotel descriptions or metadata of cameras),
but enables them to get an experiential impression of the
products. Specifically, a system could instead use
compositions of pivotal scenes captured from the movies, photos
of the hotels and their amenities, or images actually taken
with the respective cameras. Thus, users would be able to
express their taste towards more general characteristics
than just towards individual products (they may find hard to
assess or do not know about).
User Model
The quality of the user model, typically learned by means
of the user’s feedback provided during interaction with
the system, is a critical determinant for the accuracy of
today’s recommender algorithms. Model-based CF [
42
]
techniques such as Matrix Factorization (MF) [
27
] are very
prominent examples that use ratings provided by users to
efficiently generate precise recommendations. The
respective methods have been improved both by algorithmic
advances as well as by considering additional and multiple
data sources [
27
]. However, we argue that an adequate
user model should not serve only as input for the
algorithms, but might also be exploited to let users adapt the
system’s output and to increase their understanding of the
recommendation process.
Indeed, user preferences can be modeled based on other
in-puts than item ratings. In principle, all forms of implicit or
explicit feedback [
22
], also given for item-tags [
43
],
contentrelated properties, etc., can be considered. In
contentbased filtering, user models are typically learned by
probabilistic methods or nearest neighbor algorithms based
on what products the user has bought, liked or viewed
before [
11
]. Even psychological aspects such as emotions or
personality can be taken into account [
39
]. However, none
of these approaches has been developed with the specific
goal of improving interactivity. In contrast, the only way to
influence the results and to (implicitly) refine the user model
is typically by giving some kind of relevance feedback [
11
].
In social RS, it has been shown that enabling the user to
adjust the importance of the mentors used for rating
prediction increases transparency and satisfaction [
17
]. But, this
is one of the only very few examples that already provide
some insights in the model by means of visualizations and
at the same time exploit it to allow the user actively
influencing the process.
Existing interactive RS, e.g. [
6, 8, 46
], are often developed
independently of model-based CF, and thus cannot
benefit from the availability of models inferred by these efficient
and accurate techniques. MF algorithms result in latent
factor models where each user is individually represented by
a vector whose entries describe how much the user is
interested in the respective factors [
27
]. While it cannot be
expected that improving the algorithms will further increase
the actual user satisfaction with the systems [
26, 40
],
latent factor models may also be used for other purposes
than generating precise recommendations. For instance,
they already have served to visualize an item landscape
by reducing the high-dimensional factor space to a
twodimensional map [
16
]. Beyond that, the information used
to model the current user’s individual interests, i.e. his or
her own user vector, may be exploited in even more
different ways. In [
38
], for example, the characteristics of an item
have been visualized by means of latent factors. Applying
the proposed method to users instead could result in
socalled 2D feature maps showing named regions that the
current user is interested in. However, the only chance for
users to affect their preference profile in model-based CF is
usually through explicit feedback given by further ratings. In
light of this fact, it is therefore-from our point of view-a major
challenge to improve these systems significantly by letting
users actively adjust the user model.
First attempts allow users to manipulate their user vector
by other means than just rating items, i.e. more directly.
With the choice-based approach mentioned before [
32
], it is
possible to navigate through the factor space to generate a
model representing the user’s situational interests.
Extending the landscape approach of [
16
] to 3D, the map’s altitude
can be used to reflect the user’s preferences (mountains
represent areas of interest while valleys indicate low
relevance) [
29
]. In addition, the user is able to reshape the
landscape in order to manipulate the user vector, thus
leading to new results. We have also investigated other ways
to import semantics into the abstract latent factor space,
particularly by associating user-provided information such
as tags with the factors [
12
]. While this was already known
to be effective in terms of objective accuracy [
27
], we have
confirmed this finding also with respect to subjective
quality [
13
]. Moreover, our approach introduces a novel way
to manipulate the latent user model by means of
easyto-understand tags. This seems especially useful in
coldstart situations, because selecting a small number of tags
leads to a meaningful new user profile without requiring
the user to rate items first. Besides, as the abstract models
are mostly opaque, hindering the user to understand the
learned profile and hence the generated recommendations,
one can imagine using the introduced semantics to better
explain the user model.
Overall, while the aforementioned approaches already
introduce more control over the user model, many more aspects
make this part of a RS particularly interesting for
increasing the level of interactivity. For example, privacy concerns
suggest that users should be able to select themselves the
information that will be stored in the user model and
subsequently exploited for generating recommendations. Since
mediating user models, i.e. importing and integrating them
from other systems [
4
], seems promising for increasing
accuracy and providing cross-domain recommendations, this
should also be considered as an important subject when
trying to bring more interactivity and transparency into a
RS.
External Context Model
Regarding long-term interests, RS are already able to
sufficiently derive the user’s preferences, learn an adequate
user model, and present him or her with well-fitting
recommendations [
26, 40, 42
]. However, the user’s context,
i.e. date and time, season, weather, location, company of
other people, used device, and many other aspects that
depend on the user’s current situation are often not
considered in the recommendation process, although a
number of context-aware recommending approaches has been
proposed in recent years [
1
]. In fact, many systems do not
even distinguish between long-term and short-term
preferences, and especially disregard that the latter are strongly
coupled with context [
15
].
A typical example is that a user might be interested in
different things depending on, e.g., the currently used device:
When using a smartphone on the go, he or she potentially
wants suggestions for open restaurants nearby, while
information that is more general would be appropriate when
sit-ting in front of a desktop PC. Such variables indicated
by the user’s external context have already been taken into
account, resulting in, among others, restaurant and travel
recommenders, music recommenders specialized for
different purposes (in the car, at the gym, for groups, etc.), or
news RS [
1
]. The advent of smartphones has increased
the research community’s interest in developing “mobile”
context-aware recommenders even more. However,
although it would be particularly useful due to their increased
complexity and since more information, i.e. context, has to
be considered, context-aware RS often lack richer
interaction possibilities [
1
].
So far, most work has been done on the algorithmic side,
either by specializing existing methods to also integrate
context or by developing techniques specifically for that use
case. More details on how to incorporate contextual
information may be found in [
1
]. However, only little attention
has been paid to increasing user control in context-aware
recommenders [
9
]. Some conversational systems adapt
their dialogues implicitly based on the user’s interaction
sequences [
33
]. Similarly, changes in the user’s interests
can be captured to fit the results [
19
]. Based on the user’s
feedback, not only the user model, but also contextual
factors can be re-fined, e.g., to filter out those restaurants that
do not suit the current situation [
1
]. Yet overall, existing
research often tries to derive the required contextual
information automatically [
1
]. While this indeed has its benefits,
letting the user actively adjust these factors is thus typically
not possible-although it would give him or her the control
which kind of information, e.g. about restaurants (nearby
and open vs. more general), is actually desired. In [
3
],
contextual information is used to explain recommendations,
for instance, by stating that a location is especially worth a
visit at a specific time of the day. In addition, the proposed
system is one of the few exceptions that allows the user to
influence which contextual factors to consider in the
recommendation process, although this is limited to switching
them on or off. Thus, finding new ways of integrating this
part of a RS with interactive control seems to be a particular
fruitful area of future research.
Presentation
The presentation of recommended items has also received
relatively little attention by comparison. Aspects such as
what information to present, how to present it, when and
how often to present it, and how much of it to present for
any given recommendation are important when discussing
inter-activity in RS. Prior work has explored the
persuasiveness of different types of recommendation lists and
combinations of text with images [
36
]. Other researchers studied
different approaches to visualize the results [
45
], suggested
a model for timing recommendations [
10
], or determined
the number of results that leads to high choice satisfaction
without increasing choice difficulty [
5
]. However, most of
this work stops short of considering interactivity a major
factor. Consequently, ways to increase user interaction at this
stage of the recommendation process remain rather
unexplored.
Our work takes into consideration the recently made
argument that novel approaches in RS can also stem from
under-standing how people make choices [
21
].
Therefore, we aim to investigate choice support strategies that
are not typically related to recommendation technologies,
such as “combine and compute” (i.e. derive relationships
from available data to show more relevant information)
and “design the domain” (i.e. adapt the interface to
facilitate choice) [
21
]. As an ex-ample, consider tourists
looking for a hotel room on a booking website. Based on the
choices they make during their search-destination,
number of nights, desired amenities, purpose of travel, etc.-the
output could be personalized not only in terms of the
recommended items, but also tailored specifically to support
the user’s needs. Stating a preference for “fitness center”
could lead to information such as opening hours, available
machines, and pricing information being displayed more
prominently, or even further content being embedded, e.g. a
map with related workout options nearby.
In general, a RS should be able to select the features most
important for adequately personalizing the presentation
ac-cording to the user’s interests and his or her situation.
There-fore, the system might also leverage the wealth of
information contained in user-generated data (i.e. reviews,
comments, tags, or individual ratings for hotel and room
characteristics) to present more relevant details about the
recommended items. To illustrate this point, consider
someone who is interested in venues that offer good Wi-Fi
connectivity. When browsing the results, he or she might find
it useful to read reviews that specifically mention aspects
such as connection speed and signal strength or that give
an overall quality assessment. To facilitate comparison, this
information could be presented in form of a graphical scale
depicting the proportion of people who rated the internet
connection positively versus those who rated it negatively.
Since people usually have more than one requirement, a
RS that can identify the most interesting attributes for the
user could enhance recommendations with such
personalized summaries, thereby increasing their trustworthiness.
The presentation of results could also be improved by using
social media data: By mining users’ past bookings as well
as their reviews, a complex network consisting of users,
hotels, and hotel attributes can be created. This would allow
identifying with greater accuracy items a user is likely to find
at-tractive based on the attributes mentioned in his or her
re-views as well as in reviews of similar users [
35
]. In
addition, the system could also extract and present, for each
recommended item, the experiences of other people who
are interested in the same attributes as the current user.
Such a net-work of “co-staying in hotels” could thus
introduce a novel way of increasing the interaction with RS.
Overall, as the issues mentioned before suggest,
recommendations often lack transparency, and are therefore
considered less trustworthy or not meeting the user’s
situational needs [
26, 40
]. Thus, we argue that also their
presentation should be adapted to better suit the current user,
for example by presenting customized summaries of the
recommended items as well as by identifying and selecting
those features for personalization that are most important to
him or her.