We have developed a dashboard application called \SoCConnect" for integrating social data from di erent social networking sites (e.g. Facebook, Twitter), which allows users to create personalized social and semantic contexts for their social data. Users can blend their friends across di erent social networking sites and group them in di erent ways. They can also rate friends and/or their activities as favourite, neutral or disliked. We compare the results of applying ve di erent machine learning techniques on previously rated activities and friends to generate personalized recommendations for activities that may be interesting to each user. The results show that machine learning can be usefully applied in predicting the interest level of users in their social network activities, thus helping them deal with cognitive overload. A visualization technique that has been shown to work well in previous work is applied to display personalized recommendations.
Social Networking Sites (SNSs) have changed how people communicate:
nowadays, people spend more time on SNSs than ever, and prefer communication
via SNSs over emails [
In this paper, we present an approach for recommending social activities in
a dashboard application called \SoCConnect", described in [
Section 2 presents the functionalities of SoCConnect. The approach for personalized recommendation of social networking activities is described in Section 3, followed by an experimentation in Section 4 to evaluate the performance. Related work on social data integration and recommendation is presented in Section 5. Finally, Section 6 summarizes the contributions of our work and discusses future research directions. 2
In this section, we provide a brief description about the functionalities of our dashboard application SoCConnect, the results of user studies supporting our design decisions for the functionalities, and the implementation of the system. 2.1
SoCConnect retrieves users' friends information and their activities on di erent SNSs. It provides three functional categories, \managing friends", \rating friends and activities", and \personalized recommendation of activities".
The rst functional category, \managing friends" contains two functions: blending friends and grouping friends. In most cases, there is some level of overlap between the sets of a user's friends on di erent social networking sites. Our system allows the user to merge the di erent accounts of a friend across SNSs, to create a single \integrated"(or \blended") friend account for this friend in the user's SoCConnent dashboard. The second function is to group friends. Users can put their friends, both individual SNS accounts and \integrated" accounts, into groups. This function allows users to express the context and semantics of friendships, which could be the shared characteristics, interests or activities between friends.
The second functional category, \rating friends and activities" allows users to rate friends or friends' activities as favourite or disliked. The favourite activities are bookmarked, which can be revisited more easily. By rating, users are able to specify a semantic characteristic (currently limited to postive/negative) of their relationships with their friends and express their preferences on activities that they nd more or less interesting and valuable.
The third functional category, \personalized recommendation of activities" recommends activities that may be interesting to users, making use of the previous ratings and the information about friend groups. 2.2
We conducted a user study to evalute our design decisions for the functionalities of SoCConnect. 1A total number of 16 subjects (all students) were involved in this study, distributed over both gender and major (Computer Science or NonCS). They were asked questions related to the functionalities during interviews. We provide here only the most relevant results.
9 ts 8 jce 7 ub 6 fS5 roe 43 b m2 u 1 N0 4 6 5 8
3 2 2 5 5
4 2 0
1-6 7-20 >20 Number of Friends (c) 1
1 0 1 2 3 4 >4 Number of Social Networking Sites (a) <50 -100 -150 -200 >200
Number of Friends
(b)
All subjects have frequently used more than one SNSs (see Figure 1(a)). Most of them have frequently used more than two SNSs. Most of them also have more than 50 friends in total (see Figure 1(b)). Almost a half of the subjects have at least 100 friends. While it can be argued that this mini-study involved only students, this group presents the majority of users on most SNSs. For example, users of age 18-35 represented collectively 90% of the users on Facebook in 2008.2 An e-business report from 2009 shows that 75% of the adults aged 18-25 have accounts on a SNS.3
The subjects were asked about the number of their friends who have user accounts on more than two SNSs. Only two subjects do not have such friends (see Figure 1(c)). More than a half of the subjects have at least 7 such friends. Several subjects (25% of all subjects) have more than 20 such friends. 81.25% of subjects answered that these friends were active on di erent sites and most of the friends have identical activities on these sites. 75% of subjects want to view these friends' activities in one place. These results con rm a strong need for the 1 The study was approved by the Behavioural Ethics Research board of the University of Saskatchewan 2 http://social-media-optimization.com/2008/05/social-network-user-demographics/ 3 http://emarketer.com/Article.aspx?R=1006882 function of blending friends. A signi cant majority (about 90%) of subjects have some friends who share similar interests, preferences or demographic information, or do some activities together. They want to create a group for these friends and include in groups some friends on di erent sites. These results support strongly our function of grouping friends.
In order to check if users would be willing to describe semantically their relationship and the content on their SNSs, we asked the subjects whether they want to tag friends and activities. Only half of them (54.25%) provided a positive answer. Tagging requires cognitive e ort. The subjects were not sure whether they want to spend much time on tagging. Some subjects also feel that not many updates (friends' activities) are important. They prefer to tag only important activities or friends to revisit later. Instead of tagging with a word or a phrase, it requires less e ort to mark an activity or friend as \favourite" or \disliked". Moreover, users of Twitter are familiar with this way of marking updates that they may wish to revisit later. This is why we provide the function of allowing users to add friends and activities as favourite or disliked, instead of a tag of any possible phrases.
The majority (68.75%) of the subjects said that they feel overwhelmed by the number of their friends' updates in one SNS. The number of updates will increase signi cantly when the friends' accounts across di erent SNSs are integrated by an application like SoCConnect. Thus, it is necessary to provide recommendations to help users navigate through their long list of friends' updates. 2.3
To represent the semantics of social for generating recommendations, we design a generic ontology consisting of four main classes: SNS account (SNSAccount), integrated account (person), activity, and group. \SNSAccount" represents a user account on a SNS. \Person" represents a person who holds one or more SNS accounts. \Activity" represents generic information about activities appearing on SNSs. Each activity has a type. It can be a user update, e.g. a new friend added by the user, or an update by a third party application, e.g. a game such as FarmVille (farmville.com), and Ma aWars (ma a-wars.com) or other applications for Facebook, or speci c clients (e.g. Tweetie, Twitdroid) or applications (e.g. Bit.ly) for Twitter. The activity may contain text and di erent types of media, such as pictures, videos and links. It may also have a target identifying the targeted user. \Group" represents a user-de ned group for keeping friends together. A member of a group can be a SNSAccount or a Person. 3
One common problem of social networking site is information overload4. As indicated in our user studies, most of the activities from friends are not very 4 http://www.stormdawg.com/2009/10/12/social-networking-and-informationoverload/ important or interesting. Christian Kreutz in his blog calls this problem "network overload"5.
everything in the list because as indicated in our user studies, most of the activities from friends are not very important or interesting. SocConnect aims to provide a personalized recommendation on activities to individual users according to a prediction generated using their preferences on previous social data. In this section, we will present a comparison of several machine learning techniques that can be used to predict users' preferences on activities and the approach selected for visualization of the personalized recommendations. 3.1
Users directly express their preferences on activities by using the function of
rating activities as favourite or disliked activities. Based on the ratings,
SocConnect can learn users' preferences and predict whether they will be interested
in new similar activities from friends. Machine learning techniques are often used
for learning and prediction. SocConnect applies the classic techniques of
Decision Trees, Support Vector Machine [
To work with the above learning techniques, an activity needs to be represented by a set of features. Table 1 summarizes a list of relevant features and some of their possible values. Each activity has an actor (creator). SocConnect allows a user to add friends into a favourite or disliked list. Using these two features, we will be able to learn whether a user tends to be always interested in some particular friends' activities or activities from a particular type of friends. As discussed in Section 2.3, each activity has a type. We also take into account the sources which activities come from, such as Facebook and Twitter, since often users have a particular purpose for which they predominantly use a given SNS, e.g. Facebook for fun, Twitter for work-related updates. From this feature, we can nd out whether a user is only interested in activities from particular social networking sites source. Di erent applications used to generate those activities are also useful to consider. For example, if a user's friend plays \Ma aWars" on Facebook but this user does not, the status updates generated from the \Ma aWars" application may be annoying to the user. We leave out the textual content of activities. One reason is that many activities, such as video uploads, do not have any textual content. Another reason is that activities may contain non-Latin language characters and the current meta-data of activities cannot re ect which language the actor is using, which makes text analysis di cult and expensive.
After learning from a user-annotated list of activities from his or her friends, each of which is represented by a set of the feature values, our system is able to predict whether a new activity from a friend will be considered as \favourite", \neutral" or \disliked" by the user. We assign an approximate weight to the new activity as follows:
8 0:5 w = < 0 : 0:5 if predicted as favourite; if predicted as neutral; if predicted as disliked. (1)
These predictions are based on the features of each activity. The next section presents how the social context, expressed by the user by grouping friends in SocConnect, in uences the recommendations. 3.2
As described earlier, SocConnect allows users to create groups and add friends into the groups. A group implies the existence of some commonalities among the members of the group or some activities that group members have been doing together. The group information provides an indirect indication about users' preferences on activities. For example, if many activities of members in a given groups are considered as favourite by a user, the activities of the other friends classi ed by the user in this group will also be likely interesting to the user. Based on this heuristic, we extend the results of machine learning, by adjusting the weight of an activity. More speci cally, for a friend in a group, if the number of favourite activities of other group members is larger than that of disliked activities, the weight of each activity from this friend will be increased. Otherwise, the weight will be decreased. Formally, suppose that the number of liked (marked as \favourite") activities of other group members in the group is F , and the number of disliked activities from them is D, then the weight of an activity from the friend will be updated as follows: w = w + 0:5
F D F + D (2) Note that in extreme cases where every activity of the other group members is considered favourite, the weight of the friend's activity will be increased by 0.5. On another hand, if every activity of the other group members is considered disliked, the weight of the friend's activity will be decreased by 0.5. Also note that w stays the same if every activity of the other group members is considered neutral by the user (F + D = 0). For a friend who belongs to several groups, the e ect of the heuristic on the weight of the friend's activity will be averaged over these groups.
This extension brings two extra levels of user interests in activities, namely \very favourite" and \very disliked". Together, we have a range of ve levels of distinction for user interests, which has been commonly used in many popular rating systems, such as Amazon (amazon.com) and TripAdvisor (tripadvisor.com). The mapping between the interest levels of users in activities and the numerical weight for the activities is summarized in Table 2. The recommendations for the activities that the user may nd interesting are integrated in the display of the activities in the activity stream that the user views in the interface of SocConnect. Colour in a spectrum that allows people with the most common type of colour-blindness (red-green) to distinguish,6 is used to represent if an activity is recommended or unrecommended according to the predicted interest level calculated for the activity (Table 2). In this way the recommendation is unobtrusive, and can be easily ignored, but in the same time, it is intuitively clear for the user since it uses the metaphor \hot" item (displayed 6 Images can be tested for appearance with simulated colour blindness at: http://www.colblindor.com/coblis-color-blindness-simulator/ in bright orange background, yellow text and large font) and \cold" item (dark purple background, blue text and small font). The metaphor allows representing a spectrum of recommendations with a larger number of values than 5, but we have picked 5 colours to represent transitions from hot through neutral (earth colour) to cold.
We have tested a visualization of items with di erent levels of interestingness
using this metaphor with users in a study in previous work [
A 10-fold cross validation was performed on the collected data from each subject, and the average performance of the machine learning techniques over the activities of all subjects are reported in Figure 3. Although the performance di erence among these techniques is not very signi cant, support vector machine (SVM) provides the best performance, and it correctly classi es 74.1% of instances in the testing data. RBF performs the worst (70%). The performance of Naive Bayes and that of Bayesian Belief Network are about the same (around 72.6%). Decision Tree performs a little better (71.4%) than RBF.
DecisionTree RBF NaiveBayes BayesNet SVM
We also performed the validation on only 50% of collected data. More specifically, for each subject, we randomly selected 50% of collected instances. For each half of the data, we performed the same 10-fold cross validation to test the performance of the machine learning techniques. We repeated this process for 10 times to get the average performance when using only 50% of collected data. 0.76 0.74 0.72 cen 0.7 a rom 0.68 freP 0.66 0.64 0.62 0.6 % 3 n o i tca 2.8 iif s s ilfcsaeog 22..64 M a t n rce 2.2 e P 2
DecisionTree RBF NaiveBayes BayesNet SVM By looking closely into the predicted results, we found that many instances were misclassi ed by only one interest level, i.e. from \favourite" to \neutral" or from \disliked" to \neutral" and vice versa. We consider these as smaller mistakes. We summarize in Figure 4 the percentage of more serious misclassi cation from \favourite" to `disliked" and vice versa. We can see that only a very few (less than 3%) activities have been misclassi ed from \favourite" to \disliked" and vice versa. SVM consistently shows its best performance in this case. Overall, the experimental results con rm the good performance of machine learning techniques in learning social networking users' preferences on their friends' activities. SVM is particularly recommended in this context.
Results shown in Figure 5 indicate that the performance when using 50% of data is consistently lower than that when using all data for the ve machine learning techniques. This implies that the performance of personalized recommendation on social activities can be much improved when more data is collected from users, as users continuously use our system.
0.76 0.74 0.72 cen 0.7 a rom 0.68 freP 0.66 0.64 0.62 0.6
Al Data 50% Data
DecisionTree RBF NaiveBayes BayesNet SVM
Using Weka's feature selection function, we can see which features are more important for individual users. We summarize in Table 3 the number of subjects for whom each feature was the most important one in the prediction. For all users, the feature \Actor" is the most important. \Actor Type", \Activity Type" or \Application" are more important for di erent users. The source of activities (i.e. whether they come from Twitter or Facebook) turns out to be not important. This interesting di erence represents the diversity of social networking users' criteria in judging whether an activity is interesting to them, re ected in their ratings. Some users mainly care about their close friends' activities. Some users care more about the applications that generate the activities, which are usually the games they are playing. The implication is that learning the user type may be useful in personalized recommendation of activities. We leave this for future work. There have been some attempts to create personal portals that aggregate a user's accounts on di erent social networking sites, for example, the Seesmic Desktop (seesmic.com), power.com and the social web browser Flock ( ock.com). They allow the user to view her pages on di erent social networking sites in one place. In this way, the users do not have to login to many di erent sites to view the updates of their friends. However, these applications do not allow users to blend or group their friends from di erent places.7 They provide just a single-login 7 http://news.cnet.com/8301-17939 109-10109878-2.html interface in which users can switch between di erent tabs, one for each social networking site.
Bojars et al. [
In contrast, we focus mainly on developing a user-centric approach for integrating users' social data (including explicit friendship) on di erent SNSs, and that allows users to organize their social data and to create their personal contexts for the social data. We also provide personalized recommendation of friends' activities from di erent SNSs that are interesting to users.
Most recommender systems use collaborative ltering [8{10] based on the
sharing of user ratings. While many SNSs deploy algorithms based on the
analysis of social network structure to recommend new friends to the user, there
haven't been many approaches to recommend contents on SNSs. One such
approach is SoNARS. It takes a hybrid approach, combining results from
collaborative ltering and content-based algorithms [
Our work has four contributions: 1) integration of social data from di erent SNSs; 2) allowing users to de ne their personal contexts of social data, including their integrated friends who may have SNS accounts on di erent SNSs, groups of their friends who share commonalities and activities from the users' own perspective, as well as their interest level (favourite, neutral or disliked) for friends and activities; 3) personalized recommendation of activities that may be interesting to individual users; 4) suggestion of a particular machine learning method for user preferences that has the best performance among ve compared methods (SVM). A fth potential contribution is the visualization of personalized recommendations integrated in the interface for viewing the activities, once its bene ts are evaluated with users. Together, the personal dashboard application SocConnect provides users with a tool of integrating social data across di erent SNSs and with the convenience to selectively view friends' activities that are interesting to them.
For future work, next step will be to conduct user studies on the user interface to evaluate the usability of the visualization of recommendations and the appropriateness of the proposed heuristic to supplement machine learning. We are interested in exploring more deeply the relative importance of di erent features of social networking activities, to further improve the performance of personalized recommendation of activities. Other features that may be worth looking at include textual content of activities and the targeted friends of friends in activities. 7
This work has been supported by the Canada NSERC Discovery Grant of the third author.