Interest-Based vs. Social Person-Recommenders in Social Networking Platforms

Social network based approaches to person recommendations are compared to interest based approaches with the help of an empirical study on a large German social networking platform. We assess and compare the performance of different basic variants of the two approaches by precision / recall based performance with respect to reproducing known friendship relations and by an empirical questionnaire based study. In accordance to expectation, the results show that interest based person recommenders are able to produce more novel recommendations while performing less well with respect to friendship reproduction. With respect to the user’s assessment of recommendation quality all approaches perform comparably well, while combined social-interest-based variants are slightly ahead in performance. The overall results qualify those combined approaches as a good compromise.

💡 Research Summary

The paper investigates two fundamentally different approaches to person recommendation on a social networking platform: a traditional social‑network‑based method that relies on the friend‑of‑friend (FoF) structure, and an interest‑based method that derives user similarity from their activity across predefined content categories. The authors use data from Utopia.de, a German community focused on environmental sustainability, which provides a relatively sparse user profile but records a rich set of user actions (posts, comments, “worth‑living” votes) that can be assigned to eleven fixed categories (C₁…C₁₁).

For the interest‑based recommender, each user’s activity in a category i is counted (actᵢ) and normalized by the user’s total categorized activity (Pᵢ) to produce a vector A = (A₁,…,A₁₁) where Aᵢ = actᵢ / Pᵢ. Similarity between two users is then computed either by cosine similarity or Pearson correlation on these vectors. Users whose similarity exceeds a configurable threshold and who are not already friends are presented as recommendations.

The FoF baseline computes similarity as 2·|Friends(u₁) ∩ Friends(u₂)| / (|Friends(u₁)| + |Friends(u₂)|), i.e., the number of common friends normalized by the average friend count of the two users.

The experimental design employs 10‑fold cross‑validation. From the full friendship graph, 10 % of the links are held out in each fold; the remaining 90 % serve as training data for the FoF recommender, while the interest‑based models use the same underlying activity data throughout. For each user, the top‑10 candidate recommendations are generated, and the number of held‑out friendships that are correctly reproduced (true positives) is recorded. Precision, recall, and F‑measure are calculated in the usual way. Because the average number of held‑out friendships per fold is only about 2, the maximum achievable precision is limited, but this constraint applies equally to all methods.

The test set comprises 334 active users (mean friends = 11.9, mean FoF ≈ 270, mean categorized activities ≈ 88). Users with fewer than three friends, fewer than eight FoF, or fewer than three categorized actions are excluded, leaving a robust sample for evaluation.

Results show that the FoF approach excels at reproducing existing friendships (higher precision and recall) but tends to recommend users who are already within the same social circle, offering limited novelty. The pure interest‑based recommender yields lower friendship‑reproduction scores but introduces a larger proportion of previously unconnected users, thereby supporting the discovery of new contacts. A hybrid model that combines the FoF similarity score with the interest‑based similarity (e.g., weighted sum) achieves a balanced performance: it improves novelty over pure FoF while retaining better reproduction than the pure interest method.

To complement the offline metrics, the authors conducted a questionnaire with 50 test participants, asking them to rate recommendation relevance, trust, and willingness to act on the suggestions. All three approaches received generally positive feedback, but the hybrid method was rated slightly higher in perceived usefulness and trustworthiness.

Key insights include: (1) activity‑derived interest profiles can generate meaningful recommendations even for users with few or no friends, addressing the cold‑start problem for passive members; (2) pure social‑network metrics are strong at reinforcing existing ties but limited for expanding the network; (3) combining social and interest signals yields a practical compromise that enhances both relevance and novelty. The study also acknowledges limitations: using friendship reproduction as the sole ground‑truth may not fully capture the multifaceted goals of person recommendation (e.g., finding discussion partners, experts, or collaborators). Moreover, the platform’s fixed‑category tagging system prevents direct comparison with free‑form social‑tagging recommenders. Future work is suggested to incorporate multi‑objective evaluation, real‑time user feedback, and richer tagging schemes to further refine person‑recommender systems.