In context-aware recommender systems, the dependency of the user's ratings on factors that describe important aspects of the recommendation context is used to provide more relevant recommendations. Individual users may be in uenced di erently by the same set of contextual factors. By understanding this kind of dependency between the user's ratings (evaluations) and context, it is possible to identify user pro les and use them to predict precisely the user ratings for items to be recommended. In this paper, we present our methodology to identify user pro les in a corpus of ratings for music tracks. These ratings were collected in a user study, which simulated typical situations that occur while driving a car. We present the ndings derived from the data, and argue that it is feasible to distinguish di erent typologies of users from the ratings they give to music tracks in speci c contexts.
H.3.3 [Information Storage and Retrieval]: Information Search and Retrieval|Information Filtering
Recommender systems predict user ratings for items on
the basis of previous ratings for similar items or similar users
[
Presented at Searching4Fun workshop at ECIR2012. Copyright c 2012 for the individual papers by the papers’ authors. Copying permitted only for private and academic purposes. This volume is published and copyrighted by its editors.
For a recommender system, there is a major implication from this observation. If we can assess such an in uence for individual users we are able to better personalize recommendations. Beyond this, it may even be possible to group users in uenced in a similar way by certain contextual conditions. This knowledge could lead to an improved prediction of ratings for items not previously rated by the user.
With this in mind, it seems worth understanding the
inuence of context on user ratings. In previous work [
Our paper is structured as follows: In the next section we brie y present our data. Next, we introduce the mathematical tools we use to analyze the in uence of context on user ratings. In sections to follow, we present evidence that context can provoke a change the music genres preferences of the user. In the nal section, we discuss whether or not the in uence of the context on ratings can even be observed for individual users, and conclude the paper with a discussion of the results and outline our plans for future work. 2.
As described in [
M IY (X; Y ) problem a small number of tracks for each genre were proposed. 95 ratings were collected per contextual factor.
For our model of context, we relied on cognitive task analyses of car driving and considered three di erent kinds of a driver's perceptions and actions as potentially relevant: Context Factor Possible Values driving style relaxed driving, sport driving road type city, highway, serpentine landscape coast line, country side,
mountains/hills, urban sleepiness awake, sleepy tra c conditions free road, many cars, tra c jam mood active, happy, lazy, sad weather cloudy, snowing, sunny, rainy natural phenomena day time, morning, night, afternoon
Situations where more than one passenger was present were beyond the scope of our research.
For the second sample, we collected tracks with ratings on a ve star scale. The sample consists of 955 ratings ignoring any context factor and 2865 ratings taking one contextual condition into account. The ratings were given by 66 di erent users (including many who had participated in the rst study). 69 to 167 ratings were collected per contextual factor depending on the assumed relevance for the experiment (see Figure 1 and the discussion in Sect. 3).
When analyzing the dependency between contextual factors and ratings we could not make any modeling assumptions regarding the nature of the dependency. The same holds for inter-factor dependencies. Therefore, parametric models for the dependency such as linear regression are not appropriate. Instead, we had to nd a non-parametric model. In information theory, the concept of mutual information of two random variables is known exactly for this purpose: it provides means to quantify the mutual dependence of two random variables.
In our case, we can apply mutual information to quantitatively assess the di erence in the average ratings for music ignoring any in uence of context compared to the average rating taking single contextual factors into account. More formally, we de ne a random variable X for the event that users assign one of the ratings 1, 2, 3, 4, or 5 to a genre (in the rst sample) or to a track (in the second sample).
Secondly, we de ne another random variable Y for the event that one of the context factors holds in the current situation. Mutual information (M I) between X and Y is then de ned as:
M I(X; Y ) = X X P (x; y) log For X we have 2436 ratings (see Section 2 above). For each of the context factors, we collected 95 ratings. Figure 1 gives a numeric overview of the average ratings in the second data set and the impact of the single context factors on the average rating.
The results indicate that users are in uenced heavily by variable driving conditions such as their own physical condition (sleepiness) and external factors such as tra c and weather. Personal factors, such as their mood, and factor not directly related to the car driving task, such as the landscape in which users are traveling, are of minor impact.
In the next step of our analysis, we wanted to understand whether the in uence of context depends on the user preference for a music track. We hypothesized that if the user more strongly likes or dislike a track then his rating can be signi cantly in uenced by contextual factors. In order to analyze this hypothesis we grouped the data into 5 partitions for each of the 5 possible ratings a user could assign to a track. I.e. the partition 1 (\the tracks disliked without considering context") contains all tracks rated with 1 (while di erent context factors were activated), and partition 5 (\the highly preferred tracks") contains the tracks rated with 5 in any context. Again, the in uence of the context factors can be computed by measuring the mutual information and therefore the dependence between the random variable \a track is rated r without considering context" (r 2 f1; 2; 3; 4; 5g) and the random variable \context factor c is active while a track is rated r". Figure 2 shows the results of this experiment. A rst look at the numbers gives the impression that the mutual information is generally higher than in the experiment documented in Figure 1. To test this in a statistically sound way, we compared the mutual information values for each partition to those shown in Figure 1 using a t-test. The results are given in the last column. With the exception of partition 3 which groups the tracks that users did rate neutrally, for each partition the di erence is statistically signi cant (the dot stands for = 0:5, for = 0:01, for = 0:001). These ndings suggest that when users have strong positive or negative opinions for certain tracks, the conditions they experience while driving a car can in uence more their ratings for these tracks.
We also analyzed the in uence of context on the preferences for certain music genres. For this purpose, we analyzed the data coming from the rst study (see above). We formalized the user responses (POSITIVE, NEGATIVE, or NONE) as a random variable I. Given this variable, the genre G and the activated context factor C given, we can estimate the probability distribution P (IjG; C) from the rst data set and compare it to the distribution P (IjG) which does not take any context into account. For our purposes, it is again interesting to compute the mutual information for the above random variables (CjG) and (IjG). The following table presents the top-3 results for all combinations of genres and context factors: 3
From these results, we can learn two lessons. First, within a given genre, the mutual information is very high only for some factors. Evidently, these have a strong in uence on the user ratings. This outcome was not obvious before the experiment as the user preferences could have been stronger than the in uence of the driving situation. However, some of these factors in uence the ratings for (almost) all genres. We may conclude that they are strongly related to the cognitive and emotional state of a driver and therefore constitute important features of recommending music in car.
Second, as the in uence of context is evident, we may conclude that even users with strong preferences for certain
We now investigate the in uence of context on individual users. We analyze the user ratings of the four users who gave most of the ratings in our second data collection phase (see above). We show that di erent contextual factors can in uence di erent users in di erent ways. In the following tables, Mean with context (MCY) is the average rating of a user for all items rated under the assumption that the given contextual factor holds. Mean without context (MCN) is the average (of all users) rating for the same items without considering context. Di erences in these averages are compared using a t-test in order to assess whether a contextual factor actually in uences the user's ratings in a signi cant way. We indicate the statistical signi cance of the di erence between MCY and MCN with the p-value of the t-test.
We note that a recommender system can exploit the
results of our data analysis when building a prediction model
that integrates the average rating of many users for an item,
a personalized component for a particular user, and a
component for the context (see [
As can be seen in column MCN in Table 3b, this user, on average, rated the tracks in the data base higher than the others. The comparison with MCN of all users (see Table 3a) suggests that for this user many of the tracks were perceived very positively in driving situations demanding the driver's attention. In fact, driving on a highway, on a serpentine or mountain road leads to an increase of the average rating (compared to MCN for all users). On the other hand, situations that can be perceived as negative (e.g. tra c jam) provoke a decrease of the user ratings. This observation similarly holds for some other factors: lots of cars, a situation quite similar to tra c jam, or driving in morning time. Interestingly, sport driving { which stands for a consciously sportive style of driving { has negative in uence on the average ratings of this user. Hence we hypothesize that the user is a ected negatively by the tracks (mainly pop music) in situations that are likely to produce stress.
User 2: Preferences around Average with Positive Tendency towards Tracks.
In this example the user has a personal average rating similar to the other users. This phenomenon is not an ef(b) MCN versus MCY of User 1 Factor MCN happy 2.432692 serpentine 2.432692 awake 2.432692 urban 2.432692 country side 2.432692 sad 2.432692 fect of any context. The sign of the signi cant di erences between MCN and MCY in Table 4a indicate that this user likes the tracks in the corpus when he feels awake. Being sad, he would never like to listen to the tracks. In general, for this user the tra c situation (di erently from user 1) seems to play a minor role. Many signi cant di erences in his ratings can be found comparing his MCY with his noncontextualized ratings (own MCN) as well as with the rating of all the users (MCN), for personal factors such as the mood and the perception of the surrounding landscape. User 3: Preferences slightly below or on Average with Negative Tendency towards the Tracks.
In this user pro le, the factors provoking signi cant differences between MCN and MCY (see Table 5a) are mostly personal ones or factors that indirectly in uence personal attitudes or the cognitive load of the driver (i.e. road type).
As many of the tracks used for our data collection were pop songs, and on average the user assigns low ratings, we can conclude that he has a strong dislike for this kind of music. This impression is strengthened by the observation that negative emotions (such as sad) lead to even worse ratings for tracks than on average for this user.
In this user pro le, there are several highly signi cant differences between the MCN of all users and MCY (see Table 6a). In every case, the tendency is negative indicating that there are almost no situations in which tracks from the data set should be recommended to such a user. Probably this user does not like the tracks in the corpus, or he even does not like to listen to music at all while driving. The signi cance level of the di erence between the personal MCN and MCY (see Table 6b), here is slightly smaller than in the previous comparison. Moreover, there is one personal factor (awake) under which the user rated signi cantly higher. But, as there are many factors with almost identical ratings to the already low non-contextualized ratings, in most situations the items should not be recommended to this user. From this observation, we can assume that as this user dislikes tracks very strongly, it is hard to nd context factors that may change his attitude. 5.
We have presented a non-parametric approach to assess the impact of a set of contextual factors on the user ratings. Our ndings from the analysis of two data collections suggest that the perceptions and experiences during the execution of a task in uence user preferences even for non-crucial items such as music tracks to be played in a car. 5.1
We found empirical evidence that the driving situation indeed in uences the driver's preferences for music. The in uence of context may even be strong enough to modify the preference of a user for his favorite tracks.
The ndings also suggest that the cognitive load of the driver, his emotional, mental, and physical state, and current tra c conditions in uence his preferences.
These ndings are surely a ected by the set of tracks used in the study. We used this set as the reported experiments were developed within an industrial project, and the tracks were provided by the media platform of the industrial partner. It is an interesting task to collect data for other set of tracks { in a wider set of types of tracks or with a di erent specialization { and repeat the analysis.
Factor sad day time active serpentine coast line # # # # # (a) MCN of all Users versus MCY of User 3 : # # # : (b) MCN versus MCY of User 3 5.2
It is important to note the constraints and conditions of
our study design. First of all, in the web survey, we created
ctive situations that the subject should imagine. Hence,
the test persons may have overestimated the relevance of
the contextual factors on their music preferences. Hence, a
di erent study where users are actually facing certain
contextual conditions is in order. But before performing that
evaluation, our study clearly indicates that users perceive
context as important and in uential, and di erent users,
with di erent music preferences, have completely di erent
perceptions. To assess this result quantitatively, the web
survey and the described methods represent a simple way to
collect and analyze data. In fact, we exploited our results in
the implementation of a real music recommender system and
player [
Currently, we are preparing a new study with an improved experimental setup: we are merging our prototype with another application that allows to log onboard data in a car. We will equip cars of test persons with this tool and collect data in real driving situations. The logged data will allow us to detect the values of certain contextual factors from onboard information about the car and its navigation system. Furthermore, we will be able to combine this data with feedback from the users (e.g., which of the recommended tracks are played or skipped). From such a new collection of data, gained in a naturalistic setting, we will validate the ndings of our simulation study.