[R] VLouvain: Louvain Community Detection Directly on Vectors, No Graph Construction
New method eliminates O(n²) bottleneck, running directly on embeddings to process massive datasets.
Researchers have introduced VLouvain, a breakthrough algorithm that reformulates the classic Louvain community detection method to work directly on vector embeddings, completely bypassing the need to construct massive similarity graphs. Traditional approaches require computing pairwise similarities between all nodes, creating O(n²) edges that become computationally prohibitive beyond ~15,000 nodes. VLouvain instead computes degrees and modularity gains from community-level vector sums, maintaining only O(n*d) state where d is the embedding dimension. This makes it mathematically identical to standard Louvain but dramatically more scalable.
In benchmark testing on the Amazon Products dataset with 1.57 million nodes and 200-dimensional embeddings, VLouvain completed in approximately 11,300 seconds while every other tested method (including cuGraph, iGraph, GVE, and NetworKit) failed before reaching half that scale. The research also revealed a critical finding: common sparsification techniques like Top-K graph construction via FAISS produce essentially random communities, with normalized mutual information scores as low as 0.04 against the full graph. When applied to GraphRAG (retrieval-augmented generation) systems, VLouvain reduced indexing time from 3 hours to just 5.3 minutes while improving retrieval recall on the MultiHopRAG benchmark from 37.9% to 48.8%.
The algorithm represents a fundamental shift in how similarity-based clustering can be performed at scale, particularly relevant for applications like recommender systems, knowledge graph construction, and finding structure in high-dimensional data. By eliminating the graph construction bottleneck, VLouvain enables analysis of datasets orders of magnitude larger than previously possible while maintaining mathematical equivalence to the established Louvain method. The code is available on GitHub, and the paper has been accepted for EDBT 2026.
- Processes 1.57M nodes with 200-dim embeddings in ~11,300 seconds where other methods fail
- Reduces GraphRAG indexing from 3 hours to 5.3 minutes with 48.8% recall vs 37.9%
- Reveals Top-K sparsification produces random communities (NMI ~0.04 vs full graph)
Why It Matters
Enables community detection at unprecedented scale for recommender systems, GraphRAG, and data analysis without approximation.