Playing the role of weak clique property in link prediction: A friend recommendation model

An important fact in studying the link prediction is that the structural properties of networks have significant impacts on the performance of algorithms. Therefore, how to improve the performance of link prediction with the aid of structural properties of networks is an essential problem. By analyzing many real networks, we find a common structure property: nodes are preferentially linked to the nodes with the weak clique structure (abbreviated as PWCS to simplify descriptions). Based on this PWCS phenomenon, we propose a local friend recommendation (FR) index to facilitate link prediction. Our experiments show that the performance of FR index is generally better than some famous local similarity indices, such as Common Neighbor (CN) index, Adamic-Adar (AA) index and Resource Allocation (RA) index. We then explain why PWCS can give rise to the better performance of FR index in link prediction. Finally, a mixed friend recommendation index (labelled MFR) is proposed by utilizing the PWCS phenomenon, which further improves the accuracy of link prediction.

💡 Research Summary

Link prediction aims to infer missing or future connections between pairs of nodes in a network, and the structural characteristics of the underlying graph play a decisive role in the success of prediction algorithms. In this paper the authors identify a previously under‑explored structural pattern, which they call “Preferential attachment to Weak Clique Structure” (PWCS). By examining twelve real‑world networks from diverse domains (biological, social, transportation, power grids, etc.) they observe that nodes tend to connect more often to neighbors that belong to densely linked subgraphs resembling weak cliques, i.e., groups where the number of internal links is high but not necessarily forming a perfect clique.

To quantify PWCS the authors first split each existing edge into two categories: “strong‑tie” edges whose two endpoints share more than a threshold β of common neighbors, and “common” edges otherwise. The threshold β is chosen per network so that the numbers of strong‑tie and common edges are roughly equal. They then enumerate all possible three‑node connected subgraphs (seven configurations) and compute three conditional probabilities: P₁ (both incident edges are strong‑tie), P₂ (exactly one is strong‑tie), and P₃ (neither is strong‑tie). In eleven of the twelve networks P₁ > P₂ > P₃ holds, confirming that the presence of strong‑tie edges dramatically increases the likelihood that the remaining pair of nodes will also become linked. In corresponding null models (randomly rewired networks preserving degree sequences) the three probabilities are essentially equal, demonstrating that PWCS is a genuine structural property rather than a statistical artifact.

Guided by this observation the authors propose a new local similarity measure called the Friend Recommendation (FR) index. For a candidate pair (i, j) and a common neighbor ℓ, they define the probability that ℓ introduces i to j as

 f_{iℓj} = 1 / (k(ℓ) – 1 – |Γ(ℓ) ∩ Γ(j)|),

where k(ℓ) is the degree of ℓ, Γ(·) denotes the neighbor set, and the subtraction of 1 removes ℓ itself while the subtraction of the common‑neighbor count excludes those already known to j (reflecting the social intuition that a friend does not introduce you to someone you already know). Summing over all common neighbors yields

 f_{ij} = Σ_{ℓ∈Γ(i)∩Γ(j)} f_{iℓj}.

The final symmetric similarity score is

 S_{FR}(i,j) = (f_{ij} + f_{ji}) / 2.

The FR index inherits desirable properties of classic local indices: it grows with the number of common neighbors, it down‑weights high‑degree intermediaries (similar to Adamic‑Adar and Resource Allocation), and crucially it explicitly incorporates the PWCS effect by penalizing introductions that would create redundant triangles.

The authors evaluate FR on the twelve networks using a standard train‑test split (90 % of edges for training, 10 % for testing) and two performance metrics: Area Under the ROC Curve (AUC) and Precision@100 (the fraction of correctly predicted missing links among the top‑100 ranked non‑existent pairs). Across all datasets FR outperforms three well‑known local baselines—Common Neighbors (CN), Adamic‑Adar (AA), and Resource Allocation (RA). The improvement is especially pronounced in networks where PWCS is strongest (e.g., the Router, PB, and Power grids), with AUC gains exceeding 0.10 over the best baseline.

To further exploit PWCS, the authors introduce a mixed index (MFR). MFR dynamically selects between FR and RA for each candidate pair based on whether the PWCS condition (both incident edges being strong‑tie) holds; if it does, FR is used, otherwise RA is applied. This hybrid approach yields modest but consistent gains over pure FR, particularly in sparse networks with low clustering coefficients where the pure PWCS signal may be weaker.

The paper’s contributions are threefold: (1) empirical evidence that PWCS is a pervasive structural phenomenon in real networks; (2) a principled derivation of a new local similarity measure (FR) that embeds PWCS into the link‑prediction process; (3) a simple yet effective hybrid scheme (MFR) that further boosts predictive accuracy. Limitations include the reliance on a manually tuned β threshold and reduced effectiveness on networks where PWCS is absent (e.g., the Metabolic network). Future work could focus on automated β selection, integration of PWCS with other mesoscopic features such as community structure or core‑periphery organization, and extension to dynamic or multilayer networks. Overall, the study demonstrates that careful exploitation of subtle structural patterns can substantially improve the performance of lightweight, locally‑computed link‑prediction algorithms.