SLPA: Uncovering Overlapping Communities in Social Networks via A Speaker-listener Interaction Dynamic Process

Overlap is one of the characteristics of social networks, in which a person may belong to more than one social group. For this reason, discovering overlapping structures is necessary for realistic social analysis. In this paper, we present a novel, general framework to detect and analyze both individual overlapping nodes and entire communities. In this framework, nodes exchange labels according to dynamic interaction rules. A specific implementation called Speaker-listener Label Propagation Algorithm (SLPA1) demonstrates an excellent performance in identifying both overlapping nodes and overlapping communities with different degrees of diversity.

💡 Research Summary

The paper introduces the Speaker‑Listener Label Propagation Algorithm (SLPA), a novel framework for detecting overlapping communities in social networks. Traditional label‑propagation algorithms (LPA) assign a single label to each node, which limits their ability to capture the reality that individuals often belong to multiple groups. SLPA overcomes this limitation by allowing each node to maintain a memory of multiple labels and by modeling information exchange as a speaker‑listener interaction that mimics human conversation.

The algorithm proceeds in three stages. First, every node’s memory is initialized with its unique identifier as a label. Second, the dynamic evolution stage repeats for a predefined number of iterations T. In each iteration, nodes are shuffled and one node is selected as a listener. All of its neighbors act as speakers; each speaker draws a label from its own memory with probability proportional to the label’s frequency (the “speaker rule”) and sends it to the listener. The listener then applies a “listener rule”—typically selecting the most frequent label among those received—and adds that label to its memory. Updates are asynchronous, meaning that a listener can immediately benefit from the most recent changes of its neighbors. After T iterations the process stops.

The third stage is post‑processing. Each node’s memory is converted into a probability distribution over labels. Labels whose observed probability falls below a threshold r are discarded. Remaining labels define communities: nodes sharing a label and forming a connected component are grouped together. Nodes that retain multiple labels are considered overlapping nodes, and nested communities are removed to retain only maximal communities.

Complexity analysis shows that initialization costs O(n), the evolution stage costs O(T·m) (or O(T·n) for sparse graphs), and post‑processing costs O(T·n). In practice, with T set to 100, the algorithm scales almost linearly with the number of nodes, as demonstrated on synthetic networks up to 50 000 nodes.

Experimental evaluation uses the LFR benchmark (n = 5 000, average degree = 10, mixing parameter μ = 0.1–0.3, overlapping degree O_m = 2–8) and several real‑world social networks (Karate, Dolphins, Les Misérables, Polbooks, Football, Jazz, NetScience, Celeb, Email, CA‑GrQc, PGP). SLPA is compared against three well‑known overlapping community detection methods: CFinder (clique percolation), Copra (label‑propagation extension), and LFM (fitness‑function expansion).

Results indicate that SLPA consistently outperforms the baselines. For overlapping‑node identification, SLPA achieves the highest F‑score across all μ values, and its performance improves as the number of memberships per node (O_m) increases, unlike the other algorithms whose accuracy degrades. For overall community quality, the extended Normalized Mutual Information (NMI) metric shows SLPA attaining the highest scores, demonstrating robustness even when the mixing parameter rises to 0.3. In real networks, SLPA maintains high precision and recall (often above 0.7) and discovers a realistic number of overlapping nodes and communities, while avoiding the fragmentation observed in Copra and the parameter sensitivity of LFM.

Key insights include: (1) retaining a memory of past labels stabilizes label propagation and reduces fragmentation; (2) probabilistic speaker selection based on label frequency captures the intuition that frequently discussed topics spread more readily; (3) asynchronous updates allow rapid diffusion of newly popular labels; (4) the simple post‑processing threshold r provides a controllable way to prune weak community affiliations.

Limitations are noted: memory usage grows linearly with the number of iterations T, which may become significant for extremely large graphs, and the optimal threshold r must be tuned empirically. Future work could explore memory compression, adaptive stopping criteria, and automatic parameter selection.

In summary, SLPA offers a conceptually simple yet powerful mechanism for overlapping community detection, combining the efficiency of label propagation with a richer, memory‑based dynamic that faithfully reflects the overlapping nature of social structures. Its superior empirical performance and modest computational demands make it a valuable tool for researchers and practitioners analyzing complex networks.