Providing valuable recommendations in the DVR domain is quite straightforward when enough information about users and/or contents is known. In this work, we discuss the possibility of recommending future live events without knowing anything else but past user programmed recording schedules.
In order to provide a better user experience by means of focused advices (e.g., recommendation of new contents), the arisen issues can be summarized in two main categories. First, some logging activity must be done to nd common usage patterns on which identify potential users' interests. Users are not willing to o er an explicit pro le when using a DVR, thus we do have, possibly, only a set of observations on their activity. This is an important challenge for many Copyright is held by the author/owner(s). Workshop on the Practical Use of Recommender Systems, Algorithms and Technologies (PRSAT 2010), held in conjunction with RecSys 2010. September 30, 2010, Barcelona, Spain. known recommendation algorithms, that exploit user proles for increasing accuracy and take into account privacy issues as well.
Second, di erently from the Video on Demand domain, the usage of an Electronic Program Guide (EPG) is not always assured. This fact brings two consequences: (a) there is no knowledge on the content the user is recording and/or watching, and (b) there is no well de ned one-to-one correspondence between a recording and a broadcast event. This leads to the impossibility of directly recommending recordings to users.
Taking into account these considerations brings us our research question: in such a domain, is it possible to give valuable live event recommendations to users, only considering their recording activity on the DVR? Users have to be brought to contents of interest, but, di erently from other approaches, we are not using anything but collaborative ltering technique on users' activity. Thus, the main contribution of our approach is the demonstration that this can be achieved without any knowledge on what is being broadcast, neither EPGs nor content classi cations.
The task of recommending live events, such as TV shows,
has been already investigated in the past years. Proposed
methods can exploit di erent ways to collect the required
information for user pro ling, as well as can make use of
various recommendation algorithms. In particular, some
approaches, such as [
Another feature to tell apart existing methods for live events
recommendation is the recommender algorithm used. A
common approach relies on the content of the programs
broadcast and it is therefore called content-based.
Examples in this category can be found in [
A solution able to increase novelty of recommendations is
collaborative ltering, like the works in [
In this work, we focus on implicit feedbacks only and we use a collaborative ltering approach to compute recommendations. Our aim is to minimize the information required as input of the recommender system, without sacri cing the novelty. The real challenge is to be able to recommend programs to users without actually knowing anything about what is broadcast on TV, since no EPG is used (di erently from existing methods).
Faucet is a PVR integrated in a podcasting service1, which
allows the recording and further downloading of Italian TV
and Radio broadcasts [
Each registered user can x the desired settings for the recording of interest. At the end of the process, her recording is scheduled for the given time and will be further available for downloading purposes. Each recording ri, thus, is de ned as a tuple < ui; ci; ti; bi; ei; pi > with the following notation: user ui sets up a recording on channel ci, starting from time bi and ending at time ei, with a title ti and a periodicity pi (e.g., once, every Tuesday, mon-fri). In Faucet, channels and periodicity values are xed (users can choose their ci and pi from a combobox), while all other elds are completely up to the user.
After the end time expires, the recording is made available to the user for downloading. In case of periodic events, the recording step can occur an unde ned number of times. After each recording step, the respective download is made available.
In this section, we want to outline what our approach is. Given no knowledge on the broadcasts, we collect the users activity to compute what we call discrete events, to be used for recommendation purposes and top chart list building.
The extraction of meaningful information from the
unstructured amount of data contained in the dataset is essential
to de ne a set of events which map the broadcast programs.
Through the event discovery phase, we can discretize the
continuous domain of timings de ned by the recordings,
creating the basis for the application of a recommender
algorithm. The basic procedure used in the discretization was
rst introduced in [
Aggregation. As the clustering occurs periodically, this operation aims to identify newly created events characterized by the same channel and periodicity of the formerly created ones, but comparable timings. Such elements refer to the same programs and are therefore merged into unique events, whose properties are updated by taking into account the values of all the similar ones.
Collapsing. A further re nement phase is required to grant the consistency of the generated events. In fact, due to the high variability of timings, especially when a new transmission appears, events which are initially considered as non referring to the same transmission tend to slowly and independently converge to more stable timeframes. This implies the need of merging them into single events.
As a result of the processing phase, given a set of recording clustered together, each one with the same channel ci and the same periodicity pi, we compute a discrete event ei in the form of: < fuig; t; ci; b; e; pi >, where fuig is the set of users whose recordings were clustered together; t is the user generated title most frequent among users in fuig; b and e are, respectively, the starting and ending time computed as the median value of all the clustered recordings.
When future events are computed from scheduled
recordings, we are thus able to propose them to users by means of
two di erent charts: (1) a global chart returning those events
computed starting from the largest groups of recordings, i.e.,
those chosen by the largest sets of users; and (2) a user-based
recommendation list, returning a set of new events of
possible interest to each user requesting it, computed through
a similarity function over the whole population. We call
them Most Popular and Rec2 (Recordings times
Recommendations), respectively. Both charts are computed by means
of the memory based collaborative ltering approach named
k -Nearest Neighbors (kNN) [
In kNN, the weight (i.e., a measure of interest) of an element ei for an user uk can be de ned as: w(uk; ei) =
X
ua2N(uk)
r(ua; ei) c(uk; ua);
(1)
where N (uk) are the neighbors of user uk and r(ua; ei) is
equal to 1 if user ua is associated to the event ei, and 0
otherwise. The coe cient c(uk; ua) represents the neighbor's
information weight for user uk. In most of the kNN-based
algorithms [
Most Popular. The MostPopular algorithm can be de ned by means of eq. (1), assuming the number of neighbors unbounded, which implies N (uk) = U; 8uk 2 U ; and c(ua; ub) = 1; 8ua; ub 2 U , with U as the set of all users. Thus, the weight is modi ed as w(uk; ei) = Pua r(ua; ei).
After calculating the weight of all elements, they are sorted in descendant order. In the MostPopular algorithm, as the set of neighbors is independent of the user, all users receive the same recommended elements, i.e., the most popular ones.
Rec2. In order to provide personal suggestions, we have to de ne a similarity function for grouping similar users (items) from which choosing the appropriate elements to recommend. Our de nition of similarity is based only on implicit feedbacks, resulting from observing the behavior of users: if she records something, then we assume that she is interested in it; otherwise, we can not infer anything about the interest of the user for that element. We are therefore considering binary feedbacks.
Given two users u and v and the associated discrete events
Eu and Ev, we can choose the similarity metric, S(u; v),
considering several well known measures (e.g., Dice, Cosine
and Matching) [
Similarly, this approach can be adopted for the item-based similarity: two events are considered similar if the share at least a single user that is associated to both of them.
In the following, we evaluate the obtained results in the event extraction process and in the recommendation of new events to users, both in Most Popular and in Rec2.
As a remainder, we are dealing with several independent, user generated recording schedules, that we cluster together and from which we compute the discrete events. In Figure 1, a view of the distribution of the recordings is given: for each detected event, the number of recordings clustered together changes according to users' activity. As it turns out, most recordings (and, thus, most users) tend to be clustered and aggregated on very few events, while there are lots of events with very few recordings. The Most Popular algorithm exploit these inner features of the resulting discrete events to compute the top chart.
We measure how accurate is the recommendation in
predicting the elements that users would program in terms of
recall. These values are computed as the average of all users'
recall values using the top n recommended elements [
102
number of recordings
103
104
We are giving particular emphasis on the recall measure; in
fact, since we do not have explicit feedbacks regarding the
user's interest in those items which have not been
considered (i.e., not programmed, nor downloaded), precision is
not very meaningful [
First, we choose di erent similarity functions to understand whether similarity in uences the results of the user-based kNN algorithm. From Figure 2(a) it is clear that, in this case, all chosen similarity metrics show nearly the same performance.
The second step is to nd the optimal value for k. Figure 2(b) shows the results with k 2 f100; 300; 500; 700; 2000g in user-based kNN (Dice similarity), and the MostPopular recommender. We omit the values of k = f500; 700g since the results are almost equal to k = 300. Compared to the MostPopular algorithm (i.e., unbounded neighbors), a value k = 100 is not enough to outperform it, whilst for k = 2000, kNN starts to converge to it. Considering the top 10 recommended elements, we can achieve the best results for k = 300, whilst k = 500 is more suitable when taking further elements. As in most cases 10 elements are su cient for a recommendation, k = 300 o ers a good trade-o between valuable recommendations and resource consumption for building the neighborhood.
A comparison among user/item-based kNN and
MostPopular is depicted in Figure 2(c). We can observe that the
latter is clearly outperformed by the other two algorithms,
especially when more than 7 recommended items are
considered. The user-based algorithm performs slightly better
than the item-based one (more noticeable with more than
15 recommended items). In general, item-based algorithms
tend to perform better because usually the number of items
is considerably lower than the users [
In this paper we show how to recommend live events to users without any knowledge about the broadcast content 30 30% 25% lcae20% R 15% 10% 5% 0%0 0%0
Dice similarity Cosine similarity Matching similarity 25 30
15 10 top selection 20 5 (a) Comparison between similarity functions in user-based kNN (b) Recall for user-based kNN (c) Recall for kNN (k = 300) wrt MostPopular and user's likes. Recommendations can be given both globally and personally. It is important to underline that the most popular events are easier to predict since users tend to naturally focus on them, even without any speci c suggestion. On the contrary, granting a high novelty in personal recommendations is a more challenging goal due to the reduced amount of explicit information. Nevertheless, we can obtain interesting results even exploiting a simple approach as the kNN. We are currently attempting other approaches to recommendation (e.g., latent factor model) with implicit feedbacks, with the aim of improving the prediction accuracy.