Enhanced Pruning for Distributed Closeness Centrality under Multi-Packet Messaging

Identifying central nodes using closeness centrality is a critical task in analyzing large-scale complex networks, yet its decentralized computation remains challenging due to high communication overhead. Existing distributed approximation techniques, such as pruning, often fail to fully mitigate the cost of exchanging numerous data packets in large network settings. In this paper, we introduce a novel enhancement to the distributed pruning method specifically designed to overcome this communication bottleneck. Our core contribution is a technique that leverages multi-packet messaging, allowing nodes to batch and transmit larger, consolidated data blocks. This approach significantly reduces the number of exchanged messages and minimizes data loss without compromising the accuracy of the centrality estimates. We demonstrate that our multi-packet approach substantially outperforms the original pruning technique in both message efficiency (fewer overall messages) and computation time, preserving the core approximation properties of the baseline method. While we observe a manageable trade-off in increased per-node memory usage and local overhead, our findings show that this is outweighed by the gains in communication efficiency, particularly for very large networks and complex packet structures. Our work offers a more scalable and efficient solution for decentralized closeness centrality computation, promising a significant step forward for large-scale network analysis.

💡 Research Summary

The paper tackles the problem of efficiently estimating closeness centrality in large‑scale distributed networks, where the communication overhead of existing approximation methods—particularly the pruning technique introduced by Masakuna et al.—becomes a bottleneck. The authors propose an enhanced pruning algorithm that integrates multi‑packet messaging, allowing a logical message to be split into m smaller packets and transmitted using a sliding‑window Go‑Back‑N ARQ protocol. This design reduces the likelihood of packet loss, improves bandwidth utilization, and enables selective retransmission of only the corrupted fragments.

Key contributions are threefold. First, the authors formally extend the original pruning method to operate under multi‑packet communication. They introduce new variables (P(t)ij, m, ŝi) to represent the packetized messages and the reduced number of transmissions per node. Second, they provide a detailed analysis of the communication overhead introduced by multi‑packet messaging, showing that each node v_i can save at least 2 · m · d_i messages (where d_i is the node degree) by limiting its active participation to D_i – 2 iterations instead of D_i. Third, they devise a “communication overload limiting” strategy that identifies prunable nodes (leaves, triangle‑forming nodes, and receive‑only nodes) early and prevents them from sending messages in later rounds, thereby preserving approximation accuracy while dramatically cutting message traffic.

The algorithmic changes are minimal: the original pruning class is augmented with packet‑related fields, and only the initial one‑hop exchange phase uses the Go‑Back‑N scheme. Subsequent rounds follow the same pruning logic as in the baseline, but each node maintains a local set F_i,t of already‑pruned neighbors to avoid unnecessary transmissions. The authors assume FIFO, bidirectional, asynchronous links, unique node identifiers, and knowledge of immediate neighbors. They also assume that no node fails at the very start of the algorithm; failures later in the execution are handled by a neighbor‑coordination mechanism adapted from prior work.

Experimental evaluation mirrors the settings of the original paper while adding several synthetic topologies and real‑world datasets (social networks, mobile ad‑hoc simulations). Four metrics are measured: total number of messages exchanged, average per‑round execution time, closeness centrality estimation error (correlation with a ground‑truth centralized computation), and per‑node memory consumption. Results show that the multi‑packet pruning approach reduces total messages by roughly 30 % on average, cuts execution time by about 18 %, and maintains a Pearson correlation of 0.97 or higher with the exact centrality values. Memory usage increases modestly (≈ 12 % per node) due to packet buffers and the storage of pruning sets, a cost deemed acceptable for modern devices.

The paper’s significance lies in demonstrating that a carefully designed packetization layer can be seamlessly integrated into a distributed pruning framework, yielding substantial scalability gains without sacrificing accuracy. The approach is especially relevant for bandwidth‑constrained environments such as IoT deployments, vehicular networks, and large sensor clusters where high‑throughput, low‑latency communication is critical. The authors acknowledge limitations: the current design assumes static topology during the initial phase, and the analysis focuses on low error‑rate channels where Go‑Back‑N is efficient. Future work is suggested on adaptive window sizing, handling dynamic node failures, and extending the method to UDP‑based lightweight protocols. Overall, the study provides a compelling, experimentally validated solution for reducing communication overhead in distributed centrality estimation, marking a meaningful step toward practical, large‑scale network analytics.