Recommender systems (RS) based on collaborative filtering (CF) or content-based filtering (CB) have been shown to be efective means to identify items that are potentially of interest to a user, by mostly exploiting user's explicit or implicit feedback on items. Even though, these techniques achieve high accuracy in recommending, they have their own shortcomings- so hybrid solutions combining the two techniques, have emerged to overcome their disadvantages and benefit from their strengths. Another general problem can be seen in the lack of transparency of contemporary RS, where the user preference model and the recommendations that represent that model are neither explained to the current user nor the user can influence the recommendation process except for rating or re-rating (more) items. In this paper, we first enhanced the CF approach by modelling user preferences based on items' features in a complex product domain. The user-feature model is then used as an input to the user-based CF to generate recommendations and explanations. With our proposed approach, we aim to increase transparency and ofer richer interaction possibilities in current Recommender Systems- where users are allowed to express their interests in terms of features and interactively manipulate their recommendations through existing user profile and explanations.
• Information systems → Recommender systems.
Recommender systems (RS) based on collaborative filtering (CF) and
content-based filtering (CB) are widely used techniques [
However, each technique introduces some shortcomings where
the CF technique has to deal with a cold-start, scalability and the so
called "Gray Sheep"1 problems [
Current state-of-the-art approaches are already quite mature and
are often applied in a relatively straightforward recommendation
scenario by mostly relying on the user ratings of the items. The
ifltering process of such systems often assumes that the features of
an item are equally important for the user. However, in reality, the
evaluation of recommendations is a complex scenario especially for
high risk-involved domains e.g., digital cameras, where deciding
to buy a digital camera is more complex than choosing a song to
listen. In such complex domains where personal and financial risk
is associated with a product decision, users rely more on item’s
features to make a decision, where features of an item play an
important role in the user’s evaluation of an item [
The increasing complexity of recommender systems has also
created a demand for more transparent explanations and most of
the research on explaining recommendations has mostly focused
on a single source of data. Combining the user ratings with item’s
features in a hybrid manner could more clearly explain the user
interests for an item and could be more efective than the explanations
that rely only on a single source of data [
Additionally, it has been shown that only improving
recommendation process in terms of increased algorithmic accuracy does not
necessarily lead to appropriate level of user satisfaction [
To address the above mentioned issues, we implemented a hybrid approach in the complex domain of digital cameras, in which we exploit the user’s preferences of features in a collaborative filtering approach that computes the similar users based on the feature preferences rather than the item preferences. We implemented a prototype system, that showcases the possibilities of exploiting the user’s preferences of features in terms of 1) incorporating in the complex recommendation process 2) explaining the user’s interest of an item 3) interactively manipulating the recommendation process. We contribute to the state of research by addressing the following research question:
RQ 1: How can feature-based information be exploited in collaborative filtering systems for: a) Preference elicitation of users in cold-start situation b) Generating item and feature recommendations based on the user profile c) Generating explanations based on the user profile d) Manipulating recommendations through existing user profile 2
In this section, we describe our implemented feature-based collaborative filtering approach, that enhances the conventional CF approaches which relies mostly on the explicit user’s ratings on items. 2.1
Recommender systems require a dataset that provides users’ interests and preferences. In most of the cases, these preferences are in terms of the set of items and the ratings that users provide to each item. There are many available datasets that can be used to implement diferent RS algorithms. For our proposed methodology, we used the Amazon ratings dataset, provided by the university of California San Diego (UCSD)2. The structure of the dataset is defined by a set of users U = {u1, ..., un } and a set of items I = {i1, ..., im }, where riu are the items assessed, riu ∈ D (implicit or explicit), by the user u in an expression domain Du . The rating value riu ∈ D is defined on the numeric scale from 1 (strongly dislike) to 5 (strongly like).
To implement the feature-based CF approach in the domain of digital cameras, we created a matrix R by only considering the user-items-ratings data of digital cameras. As our approach focuses on exploiting features of an item, it is necessary to complete the information provided by the matrix R with information that describes the product’s content in terms of features. This descriptive featurebased information of the cameras has been obtained from a vendor independent test organization (test.de)3 in Germany and is used to create the item-feature matrix I. This matrix represents a set of items I = {i1, , ..., im } described by a set of features C = {c1, ...., ck }, where each item i is described by a vector Vi = {vci , c = 1, ..., k }.
To associate the user-item-ratings matrix R with item-features matrix I, we created another user-item-features matrix F. For each user u, the matrix F is created and is shown in the Figure 1.
Due to some technical limitations, the features of only 447 digital cameras out of 7659 cameras present in the matrix R, were extracted from test.de website, where 128 cameras out of 447 cameras have ratings present in the matrix R. This reduces our data to 15676 users, 128 items and 16071 ratings, where users who have rated three or more than three items are considered for implementing our prototype system. 2.2
To implement our feature-based collaborative filtering approach, we followed the steps that are explained in the later sections. 1) Preference elicitation at cold-start: The first step is to create a user’s 2http://jmcauley.ucsd.edu/data/amazon/links.html 3https://www.test.de/ profile based on features instead of items, to use as an input in the feature-based CF approach.
2) The user-feature-weighting matrix Q construction: Current CF
approaches exploits user’s ratings on the items to identify similar
users and then predict items from similar users’ profiles to
recommend it to the active user. The filtering process of such systems,
does not takes into account the feature preferences which is
subjective and is diferent for each user. To implement the feature-based
CF approach, feature preferences of users are required which could
be implicitly derive from the item-ratings [
3) Computing similar users: The user-feature-weightings matrix
Q from the step 2 and the active user’s profile are used as an input
in our feature-based CF approach to compute similar users based
on the feature preferences. To compute similarity between active
user’s and other users’ feature-based profiles, the Gower’s similarity
measure is used [
5) The feature recommendation generation process: The userfeature-weighting matrix is also used as an input to CF approach to generate feature recommendations based on similar users’ feature preferences.
In the following sections, we describe each step in detail. 2.2.1 Preference elicitation at cold-start. In our feature-based CF approach, new users, in contrast to a conventional preference elicitation phase, be asked to select a certain number of features, rate the features in terms of the relevance to their needs using the 5star ratings, and select the preferred feature-value(s). For numeric features, the user can only select one customized range value and rate the feature, whereas for nominal features, the user can select multiple options using check boxes. This creates a user-featureratings profile for the new user. If the user profile already exists and active user updates his/her preferences, the newly selected value for the numeric feature(s) over writes the already present value in the profile, whereas for nominal features, the newly selected value(s) is added to the user profile (if that value does not already exists in the profile). 2.2.2 The user-feature weighting matrix Q construction. 5
The feature-based preferences of users are missing in
conventional item-ratings data sets available for RS, which needs to be
derived implicitly from users’ ratings profiles. For this purpose, the
feature-weighting technique from [
Calculation of inter-user dissimilarity. In this step, the aim is to
look for features that may describes the taste and necessities of
the user. For this purpose, the entropy Hj is computed for each
feature cj , which takes into account the discriminating aspect of the
features i.e., features with more values are more discriminant than
features with fewer values. Therefore, the higher entropy value
means that the feature is more relevant. To calculate the entropy
values for each feature, Shannon entropy method is used, which not
only give importance to features with more values but also take
care of the distribution of these values, where the feature with more
uniform distribution gets he higher entropy value [
Here, ck is the feature that takes the set of values kj , fkj is the
frequency of the feature value vk in the whole set of items I. The
log(0)=0 ensures that the values with frequency 0 does not afect the
result. The entropy value Hj is normalized to get the final entropy
Hj∗ having the values between 0 to 1.
5Please refer to [
PCCu j = r Íi riuviuj − Íi riu Íi viuj
nu s Íi (riu )2 − (Íi riu )2 nu Íi viuj 2 − Íi Viuj 2 nu
V Cu j = • Dependency coeficient for nominal features using Cramer
V coeficient Cramer V coeficient is used to measure the dependency between user’s ratings on items and the values of the nominal feature j on this set of items, and is computed by the formula: v u u u u uuuu Í Í t ku kj
fku fkj 2 fku ,kj − nu fku fkj
nu numin(|Du |, |Dj |)
Calculation of features weights. The final weight for each feature for each user is computed by taking a product of normalize entropy and dependency coeficient and is given by the formula: wuj = DCu j .H j∗ 2.2.3 Computing similar users. Once the new user’s profile is created or the active user’s profile is updated, the next step is to find users that are similar to the active user in terms of the preferred feature-value(s) and the feature-weighting. To compute a similarity, the matrix Q and the profile of active user are used as an input in the user-based collaborative filtering.
As the features of cameras are heterogeneous in nature
(nominal and numeric), so separate measures needs to be applied to
compute the similarity between two objects, considering the data
types of features in to account. For this purpose, Gower’s
coeficient of similarity [
The Gower’s general similarity coeficient measures the similarity between two objects (which in this case are active user and the (2) (3) (4) other user u) based on the variable c, with constant weight wc , and is computed using the formula:
S(au,u) = Ík
c=1 s(au,u)c wc (x(au)c , x(u)c ) Ík
c=1 δ(au,u)c wc (x(au)c , x(u)c )
Where au denotes the active user, s(au,u)c denotes the contribution provided by the c-th variable between objects au and u and the coeficient δ(au, u)c determines whether the comparison can be made for the c-th variable between objects au and u. Here wc (x(au)c , x(u)c ) indicates that the weight for variable c is a function of variable values x(au)c and x(u)c for objects au and u.
In our approach, we applied the above mentioned method in three steps: 1) Firstly, the similarity coeficient s(au,u)c is computed between the feature values of two objects. , 2) Secondly, the similarity coeficient wc is computed between the feature weights of two objects, 3) In the third step, the final similarity score S(au,u) between two objects is computed by putting the values of similarity coeficients s(au,u)c and wc in the equation (5).
Similarity between the feature-values. The first step is to compute the coeficient s(au,u)ck between the active user and the other user for each value of the feature ck . As Gower’s similarity measure takes into account the type of features when computing similarity, so the method provide separate formulas to compute the similarity coeficient s(au,u)c for nominal and numeric features.
• Computing s(au,u)c for nominal features The value of s(au,u)c for nominal variable is equal to 1 if x(au)c = x(u)c (objects au and u have the same state for the attribute c) or 0 if x(au)c , x(u)c (objects au and u have diferent state for the variable c). The comparison coeficient δ(au,u)c is equal to 1 if both objects au and u have observed states for attribute c and zero otherwise.
• Computing s(au,u)c for numeric features Gower’s similarity coeficient S(au,u)c for the numeric features is defined as:
S(au,u)c = 1 − | x(au)c − x(u)c | rc where rc is the range of values for the c-th variable.
In our proposed approach, numeric features have also been considered as nominal features. The reason is the customized range value of a numeric feature that the active user au can select, rather than a distinct value. Because of the range values, the equation (6) can not be applied as it requires the distinct feature value. By considering the numeric features as nominal features, the similarity coeficient s(au,u)c gets the value 1, if the value in the user’s u profile lies within the active user’s selected range value, otherwise the value is zero.
To compute the coeficient s(au,u)ck between two users for the feature ck , the coeficient s(au,u)ck needs to be computed between the value of the feature ck for the active user, with all values of the same feature ck in the user’s u profile. The average of the coeficient values for the feature ck would give the overall contribution of the c-th variable in computing similarity between the active user and the other user. (5) (6)
Similarity between the feature-weights. The next step is to
compute the similarity coeficient w(au,u)ck , between the weightings
of features in the active user’s au and the user’s u profile. The
weightings computed for users as shown in the section 2.2.2, ranges
between [
Computing final similarity score between two objects. The final similarity coeficient S(au,u), which determines the overall similarity between the active user and the other user in terms of featurevalues and feature-weightings is computed, by putting the values of coeficients
s(au,u)ck and w(au,u)ck in equation (5). 2.2.4 The item recommendation generation process. Once the similarity score s(au, u), is computed between the active user au and all other users, the top 20 users with highest similarity scores are considered as similar users. From the nearest neighbors profiles, the top rated items are selected, and the items that have all the active user’s preferred feature-values (again determined by applying Gower’s similarity measure), are recommended to the active user. 2.2.5
The feature recommendation generation process.
Calculating the prediction scores of features for the active user. In addition to the item recommendations, the features along with the feature values are recommended to the active user, allowing the user to explore diverse features. Once the similar users are identified, the prediction score of the feature ck , which is not in active user’s profile is computed from the similar users’ profiles, using the formula: pred(au,ck ) = r au + Íu ∈N (S(au,u)) ∗ (ru,ck − ru ) Í u ∈N (S(au,u)) (8)
Here, r au is the active user’s average rating for all selected features, N represents the total number of similar users, r(u,ck ) is the user’s u rating of the feature ck , ru is the average rating of the features for the user u, and S(au,u) is the final similarity score of the active user and the other user as computed in section 3.4.2. In the above formula, the feature-ratings of other users are considered to make prediction for a rating of the active user for the feature ck .
Calculating the feature-values for the active user. Once the prediction scores are computed for each feature based on similar users’ profiles, then the top-N features are selected for recommendation. In case of nominal features, the most frequently selected value is used to recommend it to the active user. In case of numeric features, the range of values to recommend are determined by adding and subtracting the standard deviation value from the expected value.
A
In this section, we describe our prototype system which is implemented based on our feature-based CF approach mentioned in section 2, that exploits feature preferences of the active user to compute similar users. The similar users’ preferences are then used to not only recommend the items, but also recommend the features to the active user. Our implemented system, provides several interaction possibilities to users allowing them to interactively manipulate their recommendation process through existing preference profile and explanations. The screen shot of our interface is shown in Figure 2, where diferent components are marked with red alphabetical circles and are explained below.
Eliciting user’s preferences in terms of features: As mentioned in section 2.2.1, a new user profile is created based on feature preferences, which is done through filters, where users can select features from the drop down list to indicate their preferences in terms of 5-star rating, and also select preferred feature values (as shown in section (F) in Figure 2). If the user profile already exists, as it can be seen in section (A), any modification in terms of the feature ratings or values, updates the existing profile and thus is reflected immediately in terms of updated recommendations.
The items and features recommendation generation process: Once the new user indicates the preferences in terms of features or the existing user modifies his/her preferences, then this existing profile along with the user-feature-weighting matrix Q (described in section 2.2.2) are used as an input to CF approach to compute similar users based on feature-weightings and feature-values. The section (B) and (C) in Figure 2 shows the item and feature recommendations based on similar users’ feature-based profiles.
Explaining item recommendations through feature-based profiles. In our implemented system we provide feature-based explanations which shows that the recommended items generated from similar users’ preferences (see section 2.2.3 and 2.2.4) are the clear representation of the active user’s feature-based profile. The relevance of similar users with the active user in terms of the active user’s preferred features can be seen by clicking on the "similar users" link (shown in section (D) of Figure 2), which further opens the pop-up window shown in Figure 4. This relevance of similar users with the active user can further be explored for each feature of the recommended item, as shown in Figure 3.
Explaining feature recommendations through similar users’ pro
Each recommended feature (see section 2.2.5) is further exifles. plained by clicking on the "Explain" link in section (C) of Figure2, showing similar users’ preferences for that feature in terms of feature-values, and can be seen in Figure 5.
Manipulating recommendations through existing user profile. In conventional CF approach, the only way provided to users to manipulate their recommendations is by allowing them to (re)-rate item(s). However, our approach not only allows users to improve their preferences by (re)-rating features, but they can also update their long-term profile by adding new features or removing already rated features, as shown in section (A) of Figure 2. This allows the current recommendations which are generated using user’s long term preferences, be continuously adapted to user’s actual preferences in real time. Furthermore, by clicking on the "add feature" button shown in section C of Figure 2, directly adds the feature along with the feature value into the filtering list, which can be used to update the recommendation process. similar users’ and the active user’s feature preferences 4
In the current work, we extended a conventional CF approach, and exploited feature-based information in CF approach in a hybrid manner to: 1) incorporate the feature-based preferences in complex recommendation process 2) to generate more transparent explanations and, 3) interactively manipulate the recommendation process. Based on the presented feature-based CF approach, we implemented a prototype system that ofers several interaction possibilities. similar users’ feature preferences
In the future, the focus will be on improving the algorithmic accuracy of implemented techniques and explore methods to visually present the explanations in a more transparent manner, which further needs to be evaluated.