With the large number of events published all the time in event-based social networks (EBSN), it has become increasingly di cult for users to nd the events that best match their preferences. Recommender systems appear as a natural solution to this problem. However, the event recommendation scenario is quite di erent from typical recommendation domains (e.g. movies), since there is an intrinsic new item problem involved (i.e. events can not be "consumed" before their occurrence) and scarce collaborative information. Although some few works have appeared in this area, there is still lacking in the literature an extensive analysis of the di erent characteristics of EBSN data that can affect the design of event recommenders. In this paper we provide a contribution in this direction, where we investigate and discuss important features of EBSN such as sparsity, events life time, co-participation of users in events and geographic features. We also shed some light on the performance and limitations of several well known recommendation algorithms and combinations of them on real data collected from meetup.com.
How sparse is the RSVP data and how it a ects collaborativeltering algorithms? In which point of the event life time users tend to provide RSVPs? How the geographic distance between the users home and active events a ect their decision on attending these events?
Are past RSVPs usefull for predicting future RSVPs? We derive important insights from this investigation that we believe will pave the way to the design of more e cient and informed recommendation algorithms. Moreover, 3RSVP stands for the French expression \repondez s'il vous pla^t", meaning \please respond" we compare several well known recommendation algorithms and discuss their performances and limitations on real data collected from the Meetup platform, a popular EBSN that o ers large portions of event data through their API. The rest of this paper is organized as follows. In Section 2 we discuss related works. In Section 3 we present the data collection and analysis. In Section 4 compare several well known algorithms from the literature and discuss their performances and limitations. Section 5 concludes the work.
In this section we summarize the most relevant related work on event recommendation.
Minkov et al. [
In [
A recent work by Khrouf et al. [
Thus, our work is complementary to the aforementioned works, where we investigate previously unexplored features of EBSN and how they can a ect the performance of event recommendation algorithms.
Meetup is one the world's largest EBSN nowadays4. It provides an on-line environment where people can meet both on-line and face-to-face. Events of all kinds are published all the time, ranging from simple get togethers to large concerts and conferences. Moreover, large portions of data are o ered through the site on-line API5, which turns Meetup into a good test bed for investigating new event recommendation approaches.
The cities chosen for our experiments were Phoenix, Chicago and San Jose, all from USA. These cities were selected because they (i) are among the top cities in number of users and events in Meetup and (ii) are located in di erent states, which represent eventual cultural di erences and thus contribute to form a rich and diverse sample to work with. Meetup is organized in on-line groups, where every group has a physical location. To collect the data, we passed the city names as seeds and retrieved all the groups located in a radius of 100 miles from a city location returned by Meetup. Then every user, event and RSVP (i.e. user-event pairs) of those groups were retrieved. The data collected comprise the period from January, 2010 to December, 2011. Table 1 presents the characteristics of the data collected. It is worth noting the extreme sparsity of RSVPs in all cities considered.
When an event is created, users can provide RSVPs to it, i.e., provide (Yes or No) responses. We consider that a user who respond with "Yes" has a higher probability to attend the event than the user who answer "No" or provide no answer. Hence, we use this response as a proxy value to the event attendance rate, as the real attendance count is not available in Meetup. This represents a major problem for most of the collaborative ltering-based recommendation algorithms which are well known to deteriorate under severe levels of sparsity.
We consider the life time of an event as the period between its creation in Meetup and its occurrence. In Figure 2 we can see that most of the events have a life time ranging from 5 to 100 days. This means that while a small percentage of events have a very short life time (1 day), most of the events are active long enough to be discovered by the users or brought to their attention.
Here we investigate when exactly the positive RSVPs occur during the life time of events. Figure 3 shows in the x-axis the 21 rst positive RSVPs6, in chronological order, regarding all the events of the three cities considered. The y-axis ranges from 0, when the event is created, to 1, when it happens. Notice that the more positive RSVPs events receive, 6Note that approximately 95% of all events have 21 or less RSVPs the closer to the events occurrence the RSVPs are given. Although the cities investigated present small variations in this respect, they follow the same overall pattern, i.e., most of the RSVPs are provided close to the occurrence of the event. This is even more visible in the events with a life time greater than 100 days, for example, which we noticed to receive more than 80% of all positive RSVPs (among the 21 considered) in the last 20% of the their life times. This observation bears several implications to the design of e ective event recommenders. For example, after the creation of the event there will be scarce collaborative (in terms of RSVPs) information to be used, leaving room to content-based approaches. As the occurrence of the event approaches, more RSVPs are provided which favours collaborative ltering-based methods and hybrid approaches.
The collaborative aspect of the data was investigated by the distribution of events co-participation by two di erent users (in terms of positive RSVPs). Our analysis suggests that approximately only 30% of the users co-participated in two or more events in all cities considered. This observation represents an empirical bound to the e ectiveness of collaborative ltering-based recommenders.
Figure 4 depicts the distance distribution between the users
home and events locations, also investigated by other works [
In this section we compare some well known top-n recommendation algorithms for the event recommendation task. We also evaluate the algorithms in di erent levels of sparsity in order to investigate their limitations.
The data sets of each city were time split in order to resemble a real world setting. We selected 6 time stamps, equally spaced in time, for splitting training and test. For each partition time stamp, we used the previous 6 months for training and the events created during these 6 months but occurring after the partition time stamp for test. The average number of users, events and user-events pairs (RSVPs) after these partitions are displayed in Table 2 .
Here we investigate the sparsity of the recommendable events, in all partitions, in the following levels: f0; 1; 2; 3; 4; 5; 6 where each level denotes the number of positive RSVPs received per event. Figure 5 shows the event sparsity level plot. The y-axis counts the number of events in the test set that has the given sparsity level in the train. This plot tell us that regardless of when we partition the data set, there will be always a large number of events with no RSVPs. Therefore, cold-start appear as an inherent problem of the event recommendation domain.
In this paper we are considering top-n item recommendations, which are usually related to the generation of a personalized ranking recommendation list. In our case, the task of the recommender is to correctly predict which events a given test user will provide positive RSVPs in the future (test set). We have used the well known Normalized Discounted Cumulative Gain (NDCG) metric truncated to 20
DCG@20(u) IDCG@20(u) (1) (2) In Equation 1 above, reli is 1 or 0 if the event at position i is relevant or not respectively, and the function IDCGp(u) returns the perfect ranking value, acting as a normalization term.
In this section we compare the following well know top-n item recommendation algorithms from the literature: Random: The recommendation list is randomly generated.
Most-Popular : The candidate events are ranked in
descending order of popularity. We de ne popularity of
an event as the number of positive RSVPs received.
Location-Aware: This algorithm ranks the events based
on their distances to the users home, assuming that
nearby events are more likely to be attended by the
user. This algorithm does not rely on RSVP data.
BPR-MF : The Bayesian Personalized Ranking [
User-KNN and Item-KNN : Correspond to the classic k-nearest neighbor collaborative ltering based on users or items. The Collaborative Analysis of Section 3.5 have an important role in these algorithms. After a grid-search, the neighborhood size was set to 100 for both algorithms.
Logistic-Regression: We also tested an hybrid algorithm where the event scores of all aforementioned algorithms (except the Random) are fed into a logistic regression model.
Figure 6 displays the recommendation performances of each algorithm in each city considered. In spite of the high sparsity levels, the KNN based algorithms attain the best performances in comparison to the other individual algorithms. One possible explanation to this result is that, in many cases, users who will attend the same event are already friends or acquaintances and therefore may have mutual in uence on the selection of future events. The Location-Aware algorithm is comparable to the Most-Popular, leading one to conclude that the geographic distance, although carrying some signal, is not among the main reasons a ecting the decision of a user in attending or not an event. Another potential reason for this result is the inaccuracy of the users home position that is approximated from its IP address. Nonetheless, since this algorithm does not rely on RSVP data, it represents a good alternative for full coldstart scenarios. Although BPR-MF is usually better than simpler KNN based recommenders in other domains, this is not the case here. This might be related to the extreme level of sparsity of EBSN, which is not observed in other papers that concentrate their experiments on denser regions of collaborative data. sparsity levels. We want to answer questions like: events having 20 or more positive RSVPs are more likely to be correctly recommended than events having 10 or less RSVPs? Figure 7 shows the results of this analysis for all algorithms in each sparsity level considered. The x-axis encode the algorithms and the colors encode the sparsity levels. As expected, the more positive RSVPs an event has, the more likely it is to be correctly recommended by all recommendation algorithms, except the Location-Aware that does not use RSVP information, the Item-KNN and the LogisticRegression that seems to deteriorate in Phoenix with the decrease of sparsity.
In this paper we approached the problem of event recommendations in EBSN. We showed that this task is more challenging than typical recommendation domains investigated by the literature since EBSN data is inherently cold-start.
One alternative to alleviate this problem is to use RSVP data, although this data is still very sparse.
We analysed important features of EBSN that can a ect the design of e ective event recommenders and compared well known algorithms on real data collected from the popular EBSN Meetup. Our main ndings are summarized below: - RSVPs tend to be given close to the occurrence of the event. - The largest majority of events are cold-start. - Despite the high sparsity of RSVP data, KNN-based algorithms appear as the best single alternative. - Matrix-factorization does not perform as well in this domain as it does in other more typical domains.
The Logistic-Regression approach is at least as good as the Item-KNN, attaining slightly better results in San Jose. Nonethe- In future work we intend to investigate the in uence of less, it is worth noticing that the overall N DCG@20 values group membership on event attendance and more sophisare very low, achieving at most 0:3 in the best cases. ticated context-aware models to exploit the contextual data of events, such as time, tags and events descriptions. The algorithms were also evaluated in terms of the event sparsity level. Here we want to investigate which events are more likely to be correctly recommended according to their