Source-linked AI summary
Studying User Footprints in Different Online Social Networks
Anshu Malhotra, Luam Totti, Wagner Meira, Ponnurangam Kumaraguru, Virgilio Almeida
TL;DR
Linking profiles across heterogeneous social networks is difficult because users can present inconsistent identities and profile information. The paper constructs public digital footprints and applies supervised classifiers to Twitter–LinkedIn profile pairs, achieving 98% accuracy, 99% precision, and 96% recall.
Problem
Inconsistent usernames and profile information across heterogeneous social networks make linking accounts belonging to the same user challenging.
Method
The paper compares six profile-field similarities in paired public accounts and trains supervised classifiers to disambiguate same-user from different-user profiles.
Results
98% accuracy, 99% precision, and 96% recall were achieved using the most promising feature and similarity-metric set.
Takeaways & Limitations
UserID and Name compared with Jaro-Winkler were the most discriminative features for cross-network profile disambiguation.
Abstract
from arXiv · showhide
With the growing popularity and usage of online social media services, people now have accounts (some times several) on multiple and diverse services like Facebook, LinkedIn, Twitter and YouTube. Publicly available information can be used to create a digital footprint of any user using these social media services. Generating such digital footprints can be very useful for personalization, profile management, detecting malicious behavior of users. A very important application of analyzing users' online digital footprints is to protect users from potential privacy and security risks arising from the huge publicly available user information. We extracted information about user identities on different social networks through Social Graph API, FriendFeed, and Profilactic; we collated our own dataset to create the digital footprints of the users. We used username, display name, description, location, profile image, and number of connections to generate the digital footprints of the user. We applied context specific techniques (e.g. Jaro Winkler similarity, Wordnet based ontologies) to measure the similarity of the user profiles on different social networks. We specifically focused on Twitter and LinkedIn. In this paper, we present the analysis and results from applying automated classifiers for disambiguating profiles belonging to the same user from different social networks. UserID and Name were found to be the most discriminative features for disambiguating user profiles. Using the most promising set of features and similarity metrics, we achieved accuracy, precision and recall of 98%, 99%, and 96%, respectively.
I. INTRODUCTION
The paper addresses the challenge of linking inconsistent, heterogeneous profiles across social networks by extracting public digital footprints and applying automated classifiers. It proposes a scalable approach that avoids authentication and network standardization while supporting privacy-related applications.
- Linking multiple online identities can support profile management, information-leakage monitoring, portability, personalization, and privacy-threat analysis.
- Users may choose unrelated usernames and provide inconsistent information across heterogeneous social networks, making profile linking challenging.
- The paper proposes a scalable automated technique that disambiguates profiles using publicly available online digital footprints.
- Automated classifiers classify input profiles as belonging to the same user or to different users.
- The approach uses publicly available data without requiring user authentication or standardization across social networks.
- The study conducts large-scale analysis of account linking across Twitter and LinkedIn and evaluates system performance in real-world settings.
II. RELATED WORK
Prior work links accounts using graphs, tags, web searches, probabilistic models, and profile attributes, but existing methods face important coverage, scalability, and matching limitations. The paper positions its approach against these constraints.
- Earlier approaches used graph-based techniques to unify accounts across social networks.
- Tag-based semantic methods achieved accuracy around 60–80%, while username web-search correlation achieved 66% accuracy.
- Previous profile-attribute work was limited to categorical and single-value text fields, excluding free text, location, and images.
- Reported limitations include dependence on particular network types or unavailable identifiers, computational expense, simple text matching, and subjective weights and thresholds.
III. USER PROFILE DISAMBIGUATION
The system represents paired profiles through feature-specific similarity scores and trains supervised classifiers to determine whether the profiles belong to the same user. Its architecture separates collection, feature extraction, classification, and evaluation.
- Each profile pair is converted into a similarity vector covering username, name, description, location, image, and connections.
- The system architecture uses an Account Correlation Extractor and Profile Crawler for collection, a Classification Engine for training, and a User Profile Disambiguator for evaluation.
- Supervised classifiers are trained on pairs from the same users and different users, then classify new profile pairs as matches or non-matches.
IV. DATASET
The dataset was assembled in two phases: identifying cross-service accounts known to belong to the same user, then crawling their publicly available profile fields. Sources included Social Graph API and social aggregators.
- Data collection first gathered true positive connections: profiles from different services known to belong to the same user.
- Social Graph API returned alternative profile URLs for declared connections, enabling retrieval of accounts belonging to the same user across services.
- The collection included 883,668 FriendFeed users and 38,755 Profilactic users through social aggregators.
- For each user, cross-service accounts were represented as an N-tuple of service identifiers, after which publicly available profile fields were crawled using unique user handles.
B. Data Summary
The dataset included 41,336 profiles from Twitter, YouTube, and Flickr, plus 29,129 Twitter–LinkedIn account pairs. Because YouTube and Flickr had many missing fields, subsequent analysis used the Twitter–LinkedIn accounts.
- 41,336 user profiles came from Twitter, YouTube, and Flickr, with each account triple belonging to the same user.
- 29,129 Twitter–LinkedIn account pairs were additionally collected.
- YouTube and Flickr had large proportions of missing profile fields.
- The study used 29,129 Twitter–LinkedIn accounts for all further analysis.
V. ONLINE DIGITAL FOOTPRINTS
The paper represents social-network profiles through comparable fields and uses similarity measures to match users across services. UserIDs may differ across networks, so Jaro–Winkler distance measures their string similarity.
- Profiles are represented as N-dimensional vectors whose dimensions correspond to fields such as username, name, location, and description.
- The study focuses on Twitter–LinkedIn matching because the services share comparable fields, including location and descriptions.
- UserIDs are unique usernames, handles, or identifiers that identify users on a social network and may differ across services.
- Profiles declared as non-English, approximately 13%, were ignored.
- Jaro–Winkler distance compares short UserID strings on a 0–1 scale, with higher scores indicating greater similarity.
B. Display name
Display names use the entered first and/or last name, while description and location fields are compared with specialized similarity metrics. Jaro–Winkler is applied to display names rather than exact matching.
- B. Display name: Display names contain the first name, last name, or both entered by the user.
- B. Display name: Jaro–Winkler distance measures display-name similarity instead of requiring exact matches.
- C. Description: Description fields are compared using tf-idf cosine similarity, Jaccard similarity, and WordNet-based Wu–Palmer similarity.
- Location comparison uses token-based substring and Jaccard scores after punctuation removal and lowercasing.
E. Profile Image
Profile images are downloaded, resized, converted to grayscale, and represented as pixel-value vectors before similarity functions are applied. Connection counts are compared with normalization or classes to account for service-specific ranges.
- E. Profile Image: Profile images are downloaded and stored locally for comparison.
- E. Profile Image: Images are resized to 48 x 48 pixels, converted to grayscale, and represented as vectors with values from 0 to 255.
- E. Profile Image: Image similarity uses Mean Square Error, Peak Signal-to-Noise Ratio, and Levenshtein measures.
- F. Number of Connections: Connection counts represent users followed on Twitter and private-network connections on LinkedIn.
- F. Number of Connections: Connection values are compared using service-wise normalization or classes because identical counts can have different meanings across services.
VI. EVALUATION EXPERIMENTS
The evaluation analyzes feature and metric usefulness for profile matching, using discriminative-capacity scores and class-value distributions across account pairs.
- 29,129 unique users formed the account-pair dataset used for analysis.
- Four scores—Information Gain, Relief, Minimum Description Length, and Gini coefficient—were used to assess feature discriminative capacity.Categorical attributes used entropy-based discretization.
- Table I reports the discriminative capacity of each feature–metric pair across the four evaluation approaches.
- Match and Non Match value distributions were compared with feature-specific box plots, omitting outliers for clarity.
B. Matching profiles
The study trains classifiers on similarity vectors for Twitter–LinkedIn account pairs and evaluates both classification and ranked candidate retrieval.
- Matching profiles: 58,258 training instances balanced positive account pairs with randomly synthesized negative pairs.Positive examples came from Social Graph account pairs; negatives paired accounts belonging to different users.
- Matching profiles: 98% accuracy, 99% precision, and 96% recall were achieved with the most promising feature and metric set.
- Matching profiles: The evaluated feature set was {namejw, useridjw, locgeo, descjaccard, imgls, connnorm}.
- Matching profiles: The real-world system queried Twitter using a LinkedIn display name, scored returned candidates, and ranked them by same-user probability.
- Matching profiles: 64% of correct profiles ranked first with all features, compared with 49% using the best-feature set.
- Matching profiles: 75% of correct profiles appeared within the first three ranked candidates.
VII. DISCUSSION
The discussion reports automated cross-network identification from public digital footprints, emphasizing discriminative profile features and real-world candidate retrieval results.
- The system identified users across Twitter and LinkedIn using public profile information, multiple features, and similarity metrics.
- UserID and Name compared with Jaro-Winkler were the most discriminative features.
- 98% accuracy, 99% precision, and 96% recall were achieved with the most promising feature and metric set.
- In real-world retrieval, the correct Twitter profile appeared among the top three results for 75% of LinkedIn users.
- The authors plan to add profile fields, generalize to other networks, and handle missing and incomplete information.