VoG: Summarizing and Understanding Large Graphs

How can we succinctly describe a million-node graph with a few simple sentences? How can we measure the “importance” of a set of discovered subgraphs in a large graph? These are exactly the problems we focus on. Our main ideas are to construct a “vocabulary” of subgraph-types that often occur in real graphs (e.g., stars, cliques, chains), and from a set of subgraphs, find the most succinct description of a graph in terms of this vocabulary. We measure success in a well-founded way by means of the Minimum Description Length (MDL) principle: a subgraph is included in the summary if it decreases the total description length of the graph. Our contributions are three-fold: (a) formulation: we provide a principled encoding scheme to choose vocabulary subgraphs; (b) algorithm: we develop \method, an efficient method to minimize the description cost, and (c) applicability: we report experimental results on multi-million-edge real graphs, including Flickr and the Notre Dame web graph.

💡 Research Summary

The paper introduces VoG (Vocabulary‑based Graph summarization), a method that compresses large graphs into a small set of interpretable substructures such as cliques, near‑cliques, bipartite cores, near‑bipartite cores, stars, and chains. The central idea is to treat graph summarization as a lossless compression problem and to apply the Minimum Description Length (MDL) principle: a model M (an ordered list of substructures) is evaluated by the total description length L(G,M) = L(M) + L(E), where L(M) encodes the model itself and L(E) encodes the error matrix E = M ⊕ A (the exclusive‑OR between the model’s adjacency approximation and the true adjacency matrix A).

Each vocabulary type has a dedicated coding scheme. Full cliques are described by the number of nodes and their IDs; near‑cliques additionally encode the presence/absence of edges using optimal prefix codes based on their density. Full bipartite cores are encoded by the sizes and IDs of the two partitions, while near‑bipartite cores also include edge‑presence codes. Stars are a special case of bipartite cores with a single hub; chains are encoded by the ordered list of node IDs, using a logarithmic cost per position. Errors are split into over‑predicted edges (E⁺) and missing edges (E⁻), each encoded separately with binomial‑based prefix codes, which allows fast local gain estimation.

Because the space of all possible models is combinatorial and NP‑hard, VoG relies on heuristics. First, candidate subgraphs are generated using one or more graph decomposition techniques (e.g., community detection, tree‑based splits). Each candidate is then labeled by the vocabulary type that yields the smallest local MDL cost, effectively turning an arbitrary subgraph into the most suitable primitive structure.

Four selection strategies are explored to assemble the final model: (i) Plain (include all candidates), (ii) Top‑10, (iii) Top‑100 (choose the candidates with the largest compression gain), and (iv) Greedy’NForget, a greedy algorithm that iteratively adds the candidate providing the greatest reduction in total description length and then “forgets” overlapping parts that become redundant. The best model among these strategies is returned.

Experimental evaluation on ten real‑world networks—including Flickr, the Notre Dame web graph, and a Wikipedia “edit‑war” graph—demonstrates that VoG achieves higher compression ratios than traditional community detection or graph‑decomposition baselines (typically 15‑30 % improvement). More importantly, the extracted structures have clear semantic meaning: stars often correspond to hub users or administrators, near‑bipartite cores reveal conflictual groups (e.g., edit wars), and near‑cliques capture tightly‑knit communities. The algorithm scales near‑linearly with the number of edges, processing graphs with several million edges in a few minutes on commodity hardware.

The contributions of the paper are threefold: (1) a principled MDL‑based formulation of graph summarization using a fixed vocabulary of intuitive substructures; (2) an efficient pipeline that combines candidate generation, MDL‑driven labeling, and heuristic model selection; (3) extensive empirical evidence that the method produces compact, interpretable summaries on large, heterogeneous networks. Limitations include the fixed vocabulary (which may miss domain‑specific patterns) and dependence on the quality of the initial candidate generation. Future work is suggested on automatic vocabulary expansion, learning‑based candidate ranking, and dynamic updating of the model for streaming graphs.