The paper presents the result on cross-representation mediation of user models in the context of movie recommendation. We analyze the possibility of initializing the user models for a content-based recommender starting from movie ratings provided by users in other social applications. We focus in particular on (i) an approach for inferring user model preferences from rating and (ii) the experimentation of several methods to determine default user-model preferences from community-based ratings. We tested di erent variations of the proposed approach exploiting a subset of the MovieLens 10M Dataset, computing rating predictions, and analyzing the mean absolute error between such predictions and the actual ratings provided by users. The results show that the approach is in general feasible, with MAEs decreasing for users who provided more ratings and stabilizing around 0.15. They also show that community-based ratings are best used by inferring from them an \average user model", whose preferences can ll in missing values from individual user models.
Nowadays not only users leave a large amount of ratings
spread among several web sites, but also the same items
(e.g., movies, books, songs, hotels, restaurants, etc.) are
rated several times in di erent systems, from e-commerce
web sites to social network systems. On the one side, if a
user is new in a target system (i.e., has few to none rated
items), but has a rating history in another source system
in the same domain, the already rated items can be
imported and used to recommend relevant items in the
target domain [
On the other side, as a practical application of knowledge
transfer in the context of a single domain (e.g. movies), a
target system may import user ratings of overlapping items
from a source system as auxiliary data, in order to address
the data-sparsity problem, which is a common challenging
problem in many newly launched collaborative ltering (CF)
recommenders, see for example [
For instance, if a company decides to launch a new movie CF recommender, it might import user ratings from another CF recommender in order to have an initial and reliable rating matrix. Such data is made available by several sources: MovieLens1 makes its ratings and its rating matrix publicly available, while theMovieDB 2 provides a large movies database rich of information regarding movies such as title, cast, images, keywords, trailers, similar movies, ratings and so on, available through a public API. In this way a source system may be used to overcome the cold start problem related to users and items,
Cantador et al. [
Information imported from a source system, for instance
a CF movie recommender, could be also used to initialize a
content-based (CB) recommender. This approach has been
de ned as (cross-representation) mediation of user models,
which is typical from CF to CB recommender systems [
In this paper we experiment with an approach that
combines some of the solutions described above. We in fact
analyze a cross-representation mediation method which, in
the context of movie recommendations. imports knowledge
on items and users from external sources, on order to avoid
asking users for information on their tastes. The knowledge
on users is in the form of user ratings, as it comes from
a CF movie recommender (in our case, MovieLens). The
user model is inferred from such ratings. In this respect,
the task is analogous to the one described in [
To summarize, this paper provides experimental results on: the exploitation of user model mediation to enable a warm start in a cross-system recommendation scenario, thus avoiding the well known cold start problem in recommender systems and the paradox of active users. the exploitation of di erent strategies for calculating default values, which are generated starting from values of the user community of the system, in order to solve the problem of missing values in the user model, which is especially problematic in the early interactions.
This paper is organized as follows: Section 2 describes our general approach and motivations; Section 3 present related work in the eld; Section 4 describes the user modeling approach, and the methods we experimented with for computing default values; Section 5 describes the experiments and their results; and Section 6 presents some discussion on future work and concludes the paper. 2.
The approach presented in this paper has been inspired by the development of ReAL CODE3 (Recommendation Agent for Local Contents in an Open Data Environment), a movie social network and a movie recommender system. ReAL CODE retrieves information from external sources, such as Facebook and TheMovieDatabase, and automatically maps it into its own knowledge base.
In particular ReAL CODE is designed to import user
data from Facebook as bootstrapping information to
initialize the user model at the beginning of the interaction,
to solve the cold start problem and to avoid the paradox of
the active user. We consider actions that users can make
on movies within Facebook: in particular we consider
ratings, expressed on a 1-5 rating scale4. Information imported
from Facebook is used to initialize the CB user model, which
however could be in principle enriched with ratings coming
from other sources as well. In ReAL CODE, user model
interest values are then updated by considering all the
possible user actions as an implicit source of interest following
an approach similar to the one described in [
The content-based user model of ReAL CODE (see Sec. 4) contains the user interest towards all the features of the \movie" object. The \movie" object is represented by means of the following ve main features: genres, whose values are taken from the genres taxonomy of TheMovieDB containing 17 values; tags, that are keywords assigned by users to the movie actors, who play roles in the movie; directors, that direct the movie; production countries, countries in which the movie has been produced and those in which has been distributed.
Notice that both the above features and the related data set are imported from TheMovieDB, which is in fact the main source for the movie knowledge base of the system.
For each imported rating (user; movie; rating), our goal is
to propagate the rating values to the features characterizing
movie in the user model of user. Sec. 4 details the
propagation method. This mediation is inspired by the work of
Berkovsky et al. [
Regarding recommendations, suggestions are given using a content-based approach, as will be described in section 5.
According to the classi cation by Cantador et al. [
ll the missing values of the user model with default values generated starting from the user community, as will be described in Section 4. Since at the beginning, the system, as usual, will miss community data, we propose to bootstrap the system with rating values coming from a source system (in particular we exploited the MovieLens data set). This solution may also help in solving the data sparsity problem. As soon as real user data become available, user data will override community and user data.
Notice that, we could also use the Facebook data for bootstrapping the system community, as example, the ratings given by the user's friends. However, from our initial testing, this approach does not completely solve the problem of missing values, since rating coming from user's friends may be extremely sparse.
In this article, we investigate the idea of user model transfer as a way to infer missing user models values for new users or reticent ones.
First, we present a study on cross-representation
mediation of user models from collaborative ltering to
contentbased between di erent systems in the same domain, in
order to address the new user problem. This issue takes place
when a user starts to use a new recommender, which has
no knowledge of the user's interests and is not therefore
able to provide recommendations. This can be addressed
by transferring user models created in another (source)
system, \translating" and importing them to the target
system. Many works exploit cross-system personalization for
this goal, such as [
User model transfer can happen in di erent ways,
depending on representation modalities: the simpler case is when
the representation modalities are the same, while the case
when representation modalities are di erent (e.g., from CF
to CB) is more complex and needs some form of
\translation" or mediation. The most favorable scenario implies that
di erent systems share user preferences of the same type
and the same type of representation. This scenario was
addressed, for example, by Berkovsky et al. with a mediation
strategy for cross-domain collaborative ltering [
An example of user model transfer from CB to CB is
provided by Wongchokprasitti et al. [
When data coming from the source system have a di erent
representation, a cross-representation mediation is necessary
[
The problem of lling in missing ratings/user's interest
values in recommender systems, both for new and reticent
users, is discussed by several authors. One possible solution
is to ll the value with community-based preferences from
another source outside the target system [
We adopt a similar approach (i.e. using community-based information), but we combine it with cross-representation mediation (i.e. the community-based information is a set of ratings from which we infer user model default preferences).
This is in line with the idea of social information access
[
As already discussed, our goal is to infer the user model for content-base recommendation from a set of movie ratings the user may have provided either within the system itself, or in other systems | possibly collaborative- ltering recommenders, collecting movie ratings as part of the user pro le.
In order to do this, we rst need to describe how each item - in our case, movie - is described. 4.1
As it is common in content-based systems, each movie m, besides being characterized by an id and a title, is further described by a set of features desc(m) = fF1; : : : ; Fng where each feature Fi is a pair (category; value).
In our experiments, we extracted movie descriptions from TheMovieDB5 and focused on the following set of categories: FCat = fgenre; directors; actors; production country; tagsg: In extracting the actor features, we limited ourselves to the ve actors identi ed as main cast. 4.2
For each user u, her user model UM(u) is constituted by two functions: the interest function intu : Features ! [0; 1] expresses the interest of user u in a certain feature F as a real number in the [0; 1] interval; the action count function actu : Features ! N counts the number of action user u performed within the system that are pertinent to a certain feature F .
Given a set of movie ratings provided by user u on a scale [l; u], we denote by mov(u) the set of movies rated by u, and by rat(u; m) the normalization on a [0; 1] scale of the true rating originally given by u. Moreover, for a given feature F , we denote by mov(ujF ) the set of movies rated by u having F in their description; formally mov(ujF ) = fm 2 mov(u) j F 2 desc(m)g.
We then compute interest and action count as follows: actu(F ) = intu(F ) = jmov(ujF )j; P mi2movies(ujF ) ri : actu(F ) (1) (2)
As an example, let us suppose F = (actor; Ewan McGregor) and that a user u has rated three movies with this F in their description:
Trainspotting Moulin Rouge! Star Wars - Episode I there are a lot of possible di erent feature values. As the number of ratings provided by the user grows larger, the number of missing values decreases; however there are bound to be missing values even for user models inferred from a large number of ratings.
In order to tackle this problem, we chose to test a communitybased approach where a default interest function intdef is computed from other users' ratings. Such function is used in place of intu whenever we need to evaluate a feature F with actu(F ) = 0.
We experimented with three methods, resulting in three di erent default interest functions:
Middle value: the external ratings are used to compute a global average rating. If ratings were given randomly in the [0; 1] interval, this value would be 0.5. However it is well known that users tend to use more frequently higher ratings than lower ones, as either they do not rate things they do not like, or - being good recommenders to themselves - they do not watch movies that are evidently not to their tastes. The default interest function intmid is in this case a constant, and it is equal to the global average rating.
Average user: the external ratings are used as if they all belonged to the same user, to infer - in the same way described above - an average user model, with its own interest function intave. This \average" interest function is then used as default.
Notoriety-based: it may be argued that, if a feature does not appear among a user ratings, it is because that user never watches - and therefore never rates - movies with that feature. So, absence from the user model may suggest a dislike. On the other hand, a feature may be absent from the user model because it is not very common and therefore the user has never encountered it, so she does not know whether she likes it or not. In order to take into account these di erent situations, we use external ratings to de ne a notoriety measure for each feature. The notoriety ntr(F ) of a feature F with category C is therefore computed as the number of ratings given to movies with F in it, normalized by the maximum notoriety achieved within C in order to obtain a value between 0 and 1. The default interest function then is computed as: intntr(F ) = intmid(F ) ntr(F ).
We denote by intu+def the interest function obtained combining a given user's interest function intu with a default interest function intdef: intu+def(F ) = (intu(F ) if actu(F ) > 0 intdef otherwise. 5.
We evaluated our approach using the MovieLens 10M Dataset
[
Therefore in our evaluation all users are regarded as \new" to our system (which has no information on them prior to user model extraction), while movies are assumed to be already described within our system (i.e. user model extraction does not have to deal with\new" movies).
We rst randomly selected 5000 users from the MovieLens DB, whose 701017 samples were used as community ratings to compute the three default interest functions intmid, intave, and intntr, according to the three di erent methods (middle value, average user, notoriety-based) described in the previous section.
We then created 10 groups of 325 users according to the number of ratings available for each of them. The rst group contained users with 1 to 25 samples, the second group contained users with 26 to 50 samples, and so on, up to the tenth group, containing users with 226 to 250 samples. The 325 users in each group were randomly selected among those with the proper number of samples.
For each group i = 1; : : : ; 10 we randomly split users samples into ten subsets j = 1; : : : ; 10 to be analyzed for ten-fold cross-evaluation: in fold j, the evaluation set Eij coincided with the j-th subset, while the training set Tij contained all the remaining samples. Thus for each user, 9=10 of her samples were used for the training set, while the remaining 1=10 was used for the evaluation set.
For each user in each group, we extracted from her training set a user model, as described in section 4.2. We then run a basic content-based recommender on the movies in the users' evaluation set, and obtained rating predictions. Finally, we computed the mean absolute error (MAE) between such predictions and the actual ratings provided by the user herself.
In order to ensure the statistical signi cance of the observed di erences between approaches, we performed a paired t-test for each hypothesis of the form \approach X has a lower MAE than approach Y on group i". that is, as the normalized weighed sum of feature category scores scorec(u; m). Each feature category score is in turn computed as the average interest the user has for the movie features associated with category c: scorec(u; m) =
PF 2desc(m);cat(F )=c intu+def(F ) jfF 2 desc(m)jcat(F ) = cgj :
As to the weights used in formula (3), we experimented with two weight sets:
Equal weights, all categories (EQ): wc = 1 for every feature category c; Equal weights, excluding production country category (FS): wc = 0 for c = production country; wc = 1 for all other categories.
The second weight set exploits the feature selection
analysis presented in [
As an example, let us suppose we want to recommend the movie The Men Who Stare at Goats to a user u with the following pertinent interest values in her user model: (genre; comedy) (actor; Ewan McGregor) (actor; George Clooney) (production country; us) (tag; new age)
Some features are missing from the user model; therefore we need to resort to the default interest function. Let us suppose we are using the intmid function, which assigns to all features the constant value 0:706.
With the combined interest function intu+mid we can now score the ve categories: scoreproduction country scoregenre scoreactor scoredirector scoretag
We tested each combination of a default interest method (mid, ave, ntr ) and a weight set (EQ, FS). For each combination we repeated the experiment ten times, each with a di erent partition of the initial samples between training and evaluation sets (ten-fold cross evaluation). We then checked the resulting MAEs for each pair of combinations in order to test the statistical signi cance of their di erences.
For each combination, we plotted the average MAE of each user group. In all the following diagrams, MAE values are represented on the vertical axis, while the number of ratings provided by users (i.e., user groups) are represented on the horizontal axis.
Figure 1 shows the comparison, for each default interest method, of the two weight sets, while gure 2 shows the comparison, for each weight set, of the three default interest methods.
We see that with both weight sets the intave default function provides the best results ( gure 2). Also, with both weight sets the middle value method (mid ) outperforms the notoriety-based approach (ntr ). The observed di erences are statistically signi cant: for the FS diagram (a), t 3:828 corresponding to a con dence 99:9%, while for the EQ diagram (b) t 2; 908, corresponding to a con dence 99%.
On the other hand, excluding the production country category ( gure 1) does not result in signi cant changes in MAEs. While we can observe, in the case of default function intave, a consistently lower average MAE for the FS weights, the observed di erence is statistically signi cant with condence 95% (t 2:16) only for user groups with more than 125 samples. This should not be taken as a negative result with respect to feature selection, since it shows that excluding the production country category does not generally a ect the prediction accuracy, even improving it under some circumstances.
Figure 3 compares our best-performing combination (FS+ave)
with the results in [
We can see in gure 3 that our approach e ectively
contrasts the tendency to obtain worse predictions for users
with larger training sets. Although it could be expected
that using a default interest function would signi cantly
improve rating predictions (in [
However, for users with 1 to 25 samples, i.e., for the
smaller training sets, the CBFS approach in [
In this paper we discussed the results of an experimental study on cross-representation mediation of user models in the context of movie recommendation. The goal was to analyze the possibility of initializing the user models for a content-based recommender starting from movie ratings provided by users in other social applications. While previous ratings are commonly used as user model in collaborativeltering approaches, in order to determine user similarity, content-based recommenders need the user model to express an \interest function", associating with each possible feature in the domain items characterization a value expressing the interest of the user for the feature itself.
In our study we focused in particular on two points concerning content-based user models:
An algorithm to infer a content-based user model from the user's previous ratings. Such an algorithm can be exploited both to initialize the user model from the ratings she provided to other (not necessarily recommender) applications, and to maintain the user model without explicitly asking for her preferences in terms of movie features.
A comparison of di erent methods for determining a default interest function from community ratings, to be used when the user's previous ratings do not allow to infer an interest value for a certain feature. In particular we experimented with a xed default value established as the average of all community ratings, with a default user model inferred by considering the whole community as a single user, and with a mixed approach based on the notoriety of a feature.
Both the user model inference algorithm and the default interest function are particularly useful in case of system bootstrapping. The user model can be initialized importing data the same user provided to other applications. Whenever these are insu cient, default values from community ratings can be used.
We benchmarked our experimental results against the
approach discussed in [
As it could be expected, the MAE decreases when the
user model is inferred from a larger number of ratings
(the approach presented in [
The best results (lowest MAE) are obtained when the default interest function is computed as a default user model, inferred from community ratings by considering the overall community as a single user.
The feature selection results presented in [
For users with a number of previous ratings higher
than 25 (up to 250 in our analysis), our system
improves over [
These results open the door to several further
investigations. A rst line of inquiry goes in the direction of
combining our approach with [
A second study may investigate the possibility of using community ratings for ne tuning the weight of each feature category, rather than adopting the coarser approach of feature selection, where each category basically weights either 1 (selected) or 0 (discarded).
Finally, is better to remind that the system here described has been presented as it was in the context of the experiment detailed in the paper. As described, in the running system the user model is initialized getting user data from Facebook as bootstrapping information at the beginning of the interaction, while the community values are bootstrapped in the way we described in the paper. In the future, we can further investigate the bene t of using the data coming from the preferences of the user's friends, which could be mixed with the ratings coming from the bootstrapping community, since Facebook data are in general too sparse for covering all the user's model missing values. However, our system is designed to acquire data from di erent sources, both for the user model, and regarding the community values used for the system bootstrapping. Thus, in the future, we may also consider other sources of information to replace or complete the ones here described. 7.