Source-linked AI summary
Folks in Folksonomies: Social Link Prediction from Shared Metadata
Rossano Schifanella, Alain Barrat, Ciro Cattuto, Benjamin Markines, Filippo Menczer
TL;DR
The paper asks whether annotation-based semantic similarity reveals social friendship more accurately than existing evidence suggests. Using Flickr and Last.fm, it combines alignment measures with null models that preserve activity while removing local correlations, then evaluates similarity-based link prediction; the analyses detect local alignment and improve on Last.fm’s recommendations, especially for active taggers.
Problem
Microscopic evidence relating shared tag usage to social links was limited, despite social media combining tagging activity with explicit social networks.
Method
The study measures lexical and topical similarity, uses null models preserving user activity while removing socially related alignment, and evaluates annotation-based similarity for friendship prediction.
Results
Similarity measures detect local lexical and topical alignment in Flickr and Last.fm and improve social-link prediction over Last.fm recommendations, especially for active taggers.
Takeaways & Limitations
Annotation metadata can provide accurate social-link predictions, including through incremental measures that support scalable recommendation algorithms.
Takeaways & Limitations
The authors plan further validation through user studies and a suggest-friend integration in GiveALink.org; the evaluation sampling may also favor Last.fm neighbor recommendations.
Abstract
from arXiv · showhide
Web 2.0 applications have attracted a considerable amount of attention because their open-ended nature allows users to create light-weight semantic scaffolding to organize and share content. To date, the interplay of the social and semantic components of social media has been only partially explored. Here we focus on Flickr and Last.fm, two social media systems in which we can relate the tagging activity of the users with an explicit representation of their social network. We show that a substantial level of local lexical and topical alignment is observable among users who lie close to each other in the social network. We introduce a null model that preserves user activity while removing local correlations, allowing us to disentangle the actual local alignment between users from statistical effects due to the assortative mixing of user activity and centrality in the social network. This analysis suggests that users with similar topical interests are more likely to be friends, and therefore semantic similarity measures among users based solely on their annotation metadata should be predictive of social links. We test this hypothesis on the Last.fm data set, confirming that the social network constructed from semantic similarity captures actual friendship more accurately than Last.fm's suggestions based on listening patterns.
1. INTRODUCTION
This paper examines how shared tagging metadata relates to social proximity in Flickr and Last.fm, addressing limited microscopic analysis of this interplay. It measures local lexical and topical alignment and tests whether annotation-based semantic similarity predicts friendship links.
- Research gap: The paper addresses limited microscopic evidence connecting shared tag usage with social links, focusing on Flickr and Last.fm.Both systems provide substantial user tag metadata and explicit social-link representations.
- Research questions: The central question is whether users closer in the social network have more aligned tag vocabularies, and whether this alignment predicts social links.The study considers both lexical alignment through shared tags and semantic similarity derived from annotations.
- Contributions: Strong correlations occur among user activity measures, including activity and centrality mixing in the social network.The paper characterizes these mixing patterns as part of its analysis.
- Contributions: A null-model analysis finds substantial local lexical and topical alignment among socially close users in Flickr and Last.fm despite no globally shared tag vocabularies.The null models separate actual alignment from statistical effects associated with user activity and network centrality.
- Contributions: For active Last.fm users, almost all tested semantic similarity measures outperform Last.fm neighbor suggestions at predicting actual friendship relations.Scalable measures such as Maximum Information Path are among the strongest performers.
2. RELATED WORK
Prior work examined tagging overlap, network structure, browsing, groups, and interest sharing, but generally lacked direct tests linking annotation similarity to explicit friendship networks. This paper builds on those studies with pairwise social-link prediction using available social and metadata signals.
- Flickr tagging studies: Earlier Flickr research found higher average vocabulary overlap among linked users but did not assess biases or other responsible correlations.The present work addresses this limitation through explicit controls for statistical effects.
- Flickr social behavior: Related Flickr studies investigated social-network evolution, browsing patterns, and groups as coordination tools without directly establishing the paper’s annotation-based friendship prediction task.These studies provide context on network mechanisms, user activity, and tag co-usage.
- Similarity and link prediction: Other work developed node-similarity and folksonomy-based semantic measures that this paper uses as a foundation for evaluating user similarity.The prior framework applies to networks of users, tags, or resources.
- Interest-sharing studies: Santos-Neto et al. studied interest sharing in systems without explicit social networks, so they evaluated discussion-group participation rather than direct friendship.Their results on the relation between interest sharing and collaboration were mixed.
- Position of this work: This paper differs by analyzing pairs of users, directly evaluating friendship, and comparing interest-based predictions with independent system suggestions.The comparison uses explicit social links and external recommendations.
3. DATA SETS
The study assembled Flickr and Last.fm datasets through public APIs, collecting tagging metadata alongside social or neighbor relations. The Flickr corpus spans a large user population and includes two explicitly defined social networks.
- Flickr data collection: Flickr data were collected through a distributed crawl of content uploaded between January 2004 and January 2006, conducted during the first half of 2007.The crawl used public Flickr API methods and divided the interval into smaller time windows.
- Flickr dataset: The Flickr dataset contains 241,031 users with tagging information and 118,144 users with group-membership information.These counts define the available metadata populations.
- Flickr networks: Network G0 contains 118,144 users and 2,263,182 contact edges, while G1 extends G0 with neighbors of its nodes.The two networks support analyses using users with tag, group, and contact information and their network neighborhood.
- Last.fm data collection: Last.fm data were constructed through public APIs to collect neighbor and friend relations plus user annotations on songs, artists, and albums.Last.fm distinguishes friends as social contacts from neighbors recommended using music-playing histories.
- Data handling: No filters were applied during data collection.
4. DATA ANALYSIS
The analysis characterizes heterogeneous and assortative user activity, then tests whether users close in the social network exhibit lexical and topical alignment beyond activity-related statistical effects. Alignment is strongly local, and semantic similarity measures can predict friendship links.
- 4.1 Heterogeneity and correlations: Flickr user activity is highly heterogeneous across social degree, group memberships, distinct tags, and total tag assignments.The distributions are broad, and the study focuses on users who use both tags and groups.
- 4.1 Heterogeneity and correlations: Users with more social contacts generally show increasing activity across tagging, group participation, and tag-assignment measures, despite substantial fluctuations.Highly connected users with very few tags or groups also occur.
- 4.1 Heterogeneity and correlations: Neighbors display assortative mixing: the average activity of a user's neighbors increases with the user's own activity across all computed activity measures.The degree assortativity is even stronger in G1 than in G0, while large-activity estimates fluctuate because very active users are scarce.
- 4.2 Lexical and topical alignment: Similarity in shared tags and groups declines rapidly with social distance, remaining substantial through distances 2−3 but becoming close to zero for d ≥4.The pattern indicates a strongly local lexical and topical alignment effect.
- 4.2 Lexical and topical alignment: A null model preserving the social network and activity assortativity shows that original users have stronger local alignment than reshuffled metadata, especially in cosine similarity.For d < 3, original shared-tag and shared-group averages exceed the null model, while cosine-similarity distributions differ clearly between original and reshuffled data.
- 4.2 Lexical and topical alignment: On Last.fm, semantic similarity measures based on annotations outperform the system's neighbor suggestions at predicting actual friendship relations for the most active users.The evaluation considers several measures from the literature, including scalable semantic similarity measures.
5. PREDICTING SOCIAL LINKS
The paper evaluates whether semantic similarity from user annotations can predict friendship links, comparing multiple aggregation schemes and similarity measures with Last.fm recommendations. Tag-based semantic predictions outperform Last.fm most clearly for highly active users, while gains are smaller for other sampling strategies.
- 5. PREDICTING SOCIAL LINKS: Semantic similarity from annotations is evaluated as a predictor of friendship against Last.fm neighbor recommendations based on listening patterns.The evaluation uses Flickr and Last.fm metadata, but reports Last.fm results because it offers a direct recommendation baseline.
- 5. PREDICTING SOCIAL LINKS: The study compares distributional and collaborative aggregation across item- and tag-based representations, using six similarity measures including MIP.The measures include cosine, overlap, matching, Dice, Jaccard, and maximum information path, producing 24 variants.
- 5. PREDICTING SOCIAL LINKS: Similarity-ranked user pairs are evaluated as predicted friendships by varying the threshold and measuring true-positive versus false-positive rates with ROC curves and AUC.Pairs are sampled using most-active, most-connected, or random user rankings, with the sampling focused on neighbors for comparison with Last.fm recommendations.
- 5.3 Results: For most connected and randomly selected users, topical similarity remains better than the random baseline but only marginally better than Last.fm recommendations.The best predictions outperform Last.fm by at most 3–4% in AUC for most connected users and 1–5% for random users.
- 5.3 Results: All topical similarity measures outperform Last.fm recommendations for the most active users except distributional item overlap, whose improvement is not significant.For distributional measures, aggregating across items by focusing on shared tags produces better predictions.
- 5.3 Results: 37% improvement is achieved by distributional MIP based on shared tags, while collaborative matching, collaborative MIP, and overlap achieve 35%, 33%, and 32%, respectively.The tag-based collaborative results provide comparable accuracy when scalability is important.
6. CONCLUSION AND FUTURE WORK
The paper finds robust local lexical and topical alignment among socially proximate Flickr and Last.fm users, and shows that annotation-based similarity improves friendship-link prediction. It also identifies scope boundaries requiring further validation and longitudinal analysis to assess causality.
- Conclusion: The study uses Flickr and Last.fm to examine how social links interact with users’ tagging data.Both systems provide tagging data alongside explicit social links.
- Conclusion: Strong correlations between user activity, social centrality, and tagging activity require null models to separate local alignment from mixing effects.The null model preserves user activity while removing local correlations, exposing biases caused by assortativity and heterogeneity.
- Conclusion: Robust overlap measures reveal clear local lexical and topical alignment as social-network distance decreases in Flickr and Last.fm.The measures cover tag-vocabulary overlap and group-membership overlap.
- Conclusion: Semantic similarity measures improve Last.fm social-link prediction, especially for active taggers, and incremental measures support scalable recommendation algorithms.The evaluation spans many literature-based measures and aggregation schemes; maximum information path is especially accurate across settings.
- Implications: These techniques can help users find friends with similar topical interests and facilitate topical-community formation in social media.The authors frame this as improving synergies between social and semantic networks.
- Future work: Future work will validate the findings through user studies and longitudinal analysis distinguishing whether social links or shared interests come first.Planned validation includes a friend-suggestion function in GiveALink.org and temporal analysis aimed at assessing causality.