With the large number of users connected to social networks, screenname duplication is a rising problem, which leads to interference when trying to recognize users. A number of algorithms have been proposed to distinguish user profiles on one or multiple social networks. The main task in this context is to have robust features. According to the state-of-the-art approaches, features can be: content and behavioural based features, that compare content similarity between posts or behaviour similarity (timestamps between posts (behavioural), or overlapping between content (content) for example). Attribute-based features that compare profiles attributes, such as gender, age, location or image. In this paper, we tackle this problem and propose SocialMatching++ a novel approach that leverages: (1) user life events such as graduation, marriage or new job, which used to enhance the behavioural approaches (2) profile biographies, which consist in small paragraphs that users write to comprise arbitrary information about themselves. These are used to enhance the attribute approaches. To evaluate our approach, we conducted experiments on 2,263 different profiles from Facebook matched with 5,694 Twitter users, and compared them with two baseline approaches. Our results show that SocialMatching++ achieves better results compared to the baselines approaches, showing that our system successfully bridges the gap between behavioural and attribute based approaches.
The process of registration on any social networking website is accomplished in an
easy fashion, launched by the creation of a new user account. Typically, it is
necessary to provide an email and a password to obtain a valid user profile. A user
can later access his own profile and modify a diversity of settings concerning several
aspects, such as profile image, location, relationship status, social network, interests
and other information. People can register themselves on several sites, moreover, the
same user can create different accounts on a single platform by providing the same
information for each different profiles while using different email accounts. So far,
social networking sites did not setup any mechanisms yet to detect discover if two
accounts are similar and merge them. This problem is mainly originating from
databases known as Record Linkage (RL). RL is the task of identifying records
corresponding to the same entity from one or more data sources accurately, [
The research community works forward to investigate new matching mechanisms. Timestamp variations between user posts, is a widely used behavioral feature. However, this feature could be weak if the user is active in only one social network rather than the others. Similarly, profile attributes like image, location, and others, might be not updated over time.
In this paper, we tackle this problem by leveraging two novel profile-matching features: (1) user life events and (2) profile biographies. Our approach enhances the behavioral approaches and attribute ones, arguing that even if a user is not active in one social network, the potential of sharing his life events such as marriage, graduation, new job, etc, could be high. Furthermore, given life events cannot be common for two different users. Profile biographies contain key information about the user, with the advantage that it is always public and easy to be fetched. The proposed approach is tested on profiles extracted from Facebook, and linked them to their correspondent Twitter accounts. The following two sections present the most relevant approaches closest to ours. In Table 1, we mention the feature(s) used by each approach. 2.1
Attribute-based approaches use basic profile attributes such as: gender, age, location
or profile images. Goga et al. 2015 [
Raad et al. [
Bennacer et al. [
Content and behavioural based approaches
Behavioural-based approaches use comparisons between user behaviours leveraging
features such as posting rates, timestamps, or comparing content of posts.
Van Le et al. [
Liu et al. [
Zafarani et al. [
Roedler et al. [
Sha et al. [
Van Le et al. [
Jain et al. [
Jain et al. 2015 [
username, display name, location, profile image, and number of connections
Based on the aforementioned state-of-the-art analysis, we propose two features life events and profiles biographies. In addition, we show that they can enhance each other. The approach we propose, called SocialMatching++, aims at linking user profiles from Facebook to their exact profiles on Twitter by using life events and biographies as two novel matching links. 3.1
SocialMatching++ is divided into: (1) LEBL (Life Event Based Linking) that link profiles using life events, and (2) DEBL (DEscription Based Linking) that link profiles based on profile biographies. Figure 2 shows a case study of SocialMatching++ (LEBL). It presents the user interaction on two different platforms (Facebook and Twitter) over time. We observe that user could have same life event (graduation) mentioned on both social networks, even if the content and behavior does not exists on Twitter. A life event post is not frequent. This means, it occurs in a specific circumstances, hence users who are not active on all their social networks and want to keep the audience updated, usually share these events, by writing a post that describe them or by adding them to their timeline. In our work, life events were extracted from the users’ timelines on Facebook. Unlike Facebook, no formal representation of life events on Twitter is available. Consequently, we have to perform alternative mechanisms to detect them; named entity recognition is used to recognize entities inside event posts.
Profile description and biography are attributes that exist both on LinkedIn and Twitter. On LinkedIn, users can write a detailed description about themselves. However, Twitter descriptions are shorter. Recently, Facebook developed a new feature that allows users to define anything about themselves using characters. People can mention any thing inside it (hobbies, life events, biographies, etc). Users can mention many things inside like new job, hobbies, favorite food, etc. 3.1
Let be a user profile on Facebook and a user profile on Twitter. is a known entity and it consists of a username ( ) which is composed of a first name, last name, a list of life events where (
, ) and a description . For each we have to match the exact user names on Twitter . (1) is the matching function and (2) the similarity function.
( , ) = ∀ ( ( ), ( )) ( ( ), ( )) is the similarity function, and is the total number of events on Facebook. We use the vector space model (cosine similarity) because we have a sequence of tokens to compare in both life events and biographies.
= cos( ) = ∑ =1
, ( ) ( ) √∑ =1 , ( )2 √∑ =1 , ( )2 (1) (2) (3) The complete architecture of SocialMatching++ is detailed in Figure 3. It starts by retrieving screennames from Facebook, in parallel the set of life events and biographies corresponding to each user. After, we find the exact matching screennames from twitter. For each Twitter account, we search for a similar life events on their timeline and compare them to those on Facebook, as well as we do the same for biographies. Finally, we decide if two user profiles are linkable or not. Two profiles are considered to be matched, if they have exact screenname matches and if the value of similarity score is higher than a predefined threshold t. 1, >
. screenname = . screename ( , ) = {
ℎ Screennames processing
Event classifier
Similarity functions
Linked pairs Profiles extraction
Profile linker Life
Events & desc. extraction In SocialMatching++, we decide to use Facebook and Twitter websites. Facebook is the largest social networking platform across the world, followed by Twitter. In our research, we observed that only 48.9% from the user profiles extracted from Facebook own a Twitter account. Facebook permits users to create structured life events that describe a certain circumstance, these events are commonly posted on the user’s wall (timeline). Users also can post a new update (status) describing their life events. Contrariwise, Twitter does not provide any official feature through which users can update their life events. Twitter users can nevertheless post their own life events as Tweets.
Each user registered on a social network have to define a valid first name and last name, which are called screen names. People normally distinguish their names to prevent ambiguity with other users, through the modification of their usernames (a username is an id that can be accessed via the URL, e.g. facebook.com/userid). For instance, the exact screen name “Hussein Hazimeh” is available in 50 different profiles on Facebook, and each one has a different username. In plus, users differentiate themselves by adding a nickname to the original screen names (e.g. Hussein Hazimeh (PhD Student)), or to write it in two different languages. Screennames retrieval process starts by acquiring a set of screennames from the Facebook directory1, this directory contains people, pages and place names sorted alphabetically, with the URL for each entity. All screennames extracted were composed of Latin characters for both Facebook and Twitter datasets. In Figure 2, we 1 https://www.facebook.com/directory started by querying Facebook with each screenname, after this we obtained a list of exact screennames from Twitter. The maximum number of matches were 18 for a Screen names
Nick name removal
Processing
To lower case
Split spaces
Screen names
FRieg. 4. ScreennRaemes Processing single Facebook screenname. All the nicknames that are different from Latin are removed in the screennames processing phase.
After all screennames are manipulated carefully, and matched them to the correspondent screennames on Twitter, we started in retrieving the life events for each username with the date for each one. The maximum number of events extracted per entity was seven.
A set of operations must be applied on biographies before storing them. Due to the messy content exists in many free texts, many special characters might exist, and that could diminish the matching performance. For this, we clean all biographies by removing stop words, stemming the text, remove special characters and defining entities inside the text.
Named entities are extracted from both life events and biographies. A named entity can be a person, organization, location or a university. In life events, user can mention their work place or organization name for example, as well as in biographies, same entities can be shared, in addition to persons. We used the state-of-the-art techniques to annotate entities. Conditional Random Fields (CRFs) are widely used models. Given a specific biography or life event text, applying CRF we can obtain and annotate all the entities inside this text.
Life Events Querying Mechanism: a Time Window Approach Each Facebook life event consists of a set of entities and a date. We model the timeline of a Twitter user as a series of time windows, Figure 5. A single time window is a fixed interval of time when the data stream is processed for querying. Suppose that we have a Facebook life event e that has a specific date d. Our objective is to query the time window with date d and try to find if we have existing similar life events to e. The similarity of two life events represented by the function S(e1, e2). S is positive if one of the two following scenarios is occurred: (1) if the value of Function (2) is greater than threshold, or if the similarity between two life events entities’ is positive.
The first two inputs of the profile linker are: the Facebook screen name and its corresponding Twitter ids. We have two matching problems, the first one is a (1 to 1) matching problem. This exists when we have only one matched screen name on Twitter. The second one is a (1 to n) matching problem. This exists when we have more than 2 existing Twitter profiles. The matching procedure starts by comparing two biographies considering both named entities and matching score to take the matching decision. For e.g.
Married one year ago. My life Sarah
Love you Sarah In the biographies (1) and (2), the similarity score is very low. However, if we consider common entities between them, we can observe that the user writes about his wife Sarah. In this case, we decide to link the two profiles rather than ignoring them. If the description comparison returns a null result, we query each user’s Twitter timeline with the named entities extracted from Facebook in a specific interval of time, Figure 5.
In all our experiments, we used Facebook and Twitter websites. Alternative social networks like Instagram have more number of users compared to Twitter. However, Instagram allows people to share only images, which is not enough to conduct matching studies. LinkedIn contains a lot of information about users. But the platform is dedicated to and we do not have access to the user’s timeline. Hence, even if the portion of connected users between Facebook and Twitter is not big, they remain two better choices due to the richness of information tweeted and posted on the user’s timelines. We conducted tests on 2,263 different profiles having 6,630 life events and 1948 profile descriptions from Facebook matched with 5,694 Twitter users. The constructed Facebook dataset is open to the public and available for download (upon request). To implement our system we used Selenium web driver coded in Java using eclipse, and used a PC with 8GB of RAM and eight cores. The complete system code can be downloaded2.
The 2,263 profiles from Facebook were selected using the following mechanism: we select two random profiles from Facebook that have a public friend list, each one of these two profiles has more than thousand friends. We have crawled these friend lists and for each user in this list we have extracted from his profile the public life events published on his timeline and his profile description. For each event, we have extracted its content and the exact date of publishing.
In Figure 7, a chart displays the total number of exact screen names matchings from Twitter that correspond to a unique Facebook profile screen name. #of profile
Screen names from Facebook 2 https://github.com/HusseinHESSO/ProfileLinking_v1.0 The maximum number of screen names matches is 18, and only 1,022 screen names were found on Twitter.
The maximum number of life events extracted from Facebook for each user is eight events. The total number of events extracted for each class is shown in table 2. For each life event, we define the named entity.
To evaluate our system, we use precision as a metric. In addition, we compare our system to two baseline systems.
With regard to life events, we chose one state-of-the-art system that compares the
behavioural similarity between two user profiles. The system is HYDRA [
Concerning biographies, we compare our system with systems that rely on profile
attribute information, and prove that even if users do not share public attribute
information, it is possible to link these profiles using biographies. @I seek ‘fb.me’.
[
We compare the precision of our system before and after defining entities. Figures 8 and 9 show how that the precision can be enhanced after defining entities using both the LEBL and DEBL approaches. We compare also the precision of our system with the baseline approaches, in Figures 10 and 11 and we show that both of our approaches are highly precise compared to the baselines. All the results shown in the figures take into account between 10 and 50 user profiles
To compare our system with the two baseline systems, we select four random profiles
that can fail to be linked using public attribute-based approaches, and show that they
can be linked using profile descriptions. Table 2 show a set of four random profiles
linked using biographies compared with @I seek ‘fb.me’ [
DEBL
#of Twitter profiles 5 studioandrew rodwell.mupungu Yes 5 1 As we observe in Table 3, our approach DEBL performs better than baselines. We can see that the baseline approaches succeed in linking users via image comparison. However, image matching can be more challenging than text matching. Some of them can link users via location too. Location can be same for different users. Hence, biographies can play a vital role when linking users.
We also compare LEBL to the behavioural approaches. We select four random profiles, and show that information existing between users’ timelines can vary in terms of content of posts, timestamps between posts and others. However, if we query the timeline with a specific type of information (Life events) we can detect that the two profiles belong to the same user. This can be enhanced in terms of accuracy (life events cannot be the same for two different users) and lack of activity in one of the two profiles. In Table 4, we show in detail the analysis of the four different profiles.
HYDRA TW(textweets),FB(p hotos & videos)
TW(latin letters),FB(Arabic
latters) TW(Last public post
2015) FB(last public post 1
hour ago) Similar hashtags
LEBL Yes Yes Yes Yes #of TW profiles 2 1 4 6
In this work, we described SocialMatching++ a novel system that links user profiles on Facebook to their Twitter accounts, using life events (LEBL) and descriptions (DEBL). After conducting a comprehensive state-of-the-art on this research problem, we found that none of the related works has used these two features. We proved that our approach can achieve promising results. We concluded that even users do not share similar profile attribute information, such as images, locations; they can be matched using profile descriptions. Furthermore, users with different timeline behaviours and content can be matched using life events.
This work is strongly intended to match the topic of dataset profiling. The two built datasets can be later used for profiling reasons as a future work, motivating that dataset profiling in the context of social media is quite novel.
However, our work still has some limitations. Users are matched only from Facebook to Twitter, and not vice versa. Life events classification needs a more sophisticated, to detect more possible number of events on both social networks. As future work, we are working on enhancing the flexibility of our system. System users can define later the similarity functions based on their expertise, in addition to the flexibility of selecting the social network of interest. In order to test the scalability of our approach, we are working on building large-scale datasets and extend the number of social network channels (Google+ is under study).