Multilevel compression of random walks on networks reveals hierarchical organization in large integrated systems

To comprehend the hierarchical organization of large integrated systems, we introduce the hierarchical map equation, which reveals multilevel structures in networks. In this information-theoretic approach, we exploit the duality between compression and pattern detection; by compressing a description of a random walker as a proxy for real flow on a network, we find regularities in the network that induce this system-wide flow. Finding the shortest multilevel description of the random walker therefore gives us the best hierarchical clustering of the network, the optimal number of levels and modular partition at each level, with respect to the dynamics on the network. With a novel search algorithm, we extract and illustrate the rich multilevel organization of several large social and biological networks. For example, from the global air traffic network we uncover countries and continents, and from the pattern of scientific communication we reveal more than 100 scientific fields organized in four major disciplines: life sciences, physical sciences, ecology and earth sciences, and social sciences. In general, we find shallow hierarchical structures in globally interconnected systems, such as neural networks, and rich multilevel organizations in systems with highly separated regions, such as road networks.

💡 Research Summary

The paper introduces the hierarchical map equation, an information‑theoretic framework that extends the original map equation from a single‑level description of random‑walk dynamics to a multilevel compression scheme. The core insight is the duality between data compression and pattern detection: by encoding the trajectory of a random walker (used as a proxy for real flow on a network) with the shortest possible description, one automatically uncovers the regularities that generate that flow. In the hierarchical version each level of the network hierarchy possesses its own codebook; when the walker moves between levels an additional “exit” code is emitted. The total description length L is the sum of the entropy of inter‑level transitions and the entropies of the intra‑level codebooks. Minimizing L simultaneously yields the optimal number of hierarchical levels, the partition of nodes within each level, and the assignment of inter‑level transition probabilities.

To find this optimum the authors develop an iterative multilevel search algorithm. Starting from an initial partition (either random or the result of the standard map equation), the algorithm repeatedly performs three operations: (1) splitting – candidate sub‑modules are created inside a module, a new lower‑level codebook is introduced, and the change in L is evaluated; (2) merging – adjacent modules are combined into a higher‑level module, again checking the impact on L; (3) refinement – a greedy local move phase is combined with simulated annealing to escape shallow local minima. Because each operation only requires recomputing transition probabilities and entropies for the affected parts of the network, the computational cost scales linearly with the number of edges, making the method applicable to networks with millions of nodes and edges. The algorithm has been integrated into the publicly available Infomap software.

The authors apply the method to four large, heterogeneous real‑world networks:

1. Global air‑traffic network (≈3 000 airports, 200 000 routes). The hierarchy reveals three clear levels: continents at the top, countries in the middle, and major hub airports at the bottom. The partition aligns closely with geopolitical boundaries, demonstrating that the flow of passengers naturally respects these large‑scale structures.

2. Scientific citation network (≈1.2 million papers, 10 million citations). Four top‑level disciplines—life sciences, physical sciences, ecology & earth sciences, and social sciences—emerge, under which more than 100 sub‑fields are organized into a four‑level hierarchy. The method captures both the broad disciplinary segregation and the finer‑grained specialty clusters, providing a quantitative map of scientific communication.

3. Human brain connectome (≈80 000 neurons, 1 million synapses). The optimal description consists of only two to three levels, reflecting the brain’s highly integrated nature: a shallow hierarchy suffices because information flow is globally distributed rather than confined to isolated modules.

4. World road network (≈2 million intersections, 5 million roads). Here a deep hierarchy of five to six levels is uncovered, corresponding to continents, countries, states/provinces, cities, and individual streets. The multilevel structure highlights regional bottlenecks and the strong spatial segregation of traffic flows.

In benchmark comparisons against modularity maximization, spectral clustering, and other community‑detection algorithms, the hierarchical map equation consistently achieves lower description lengths (the objective function) and higher normalized mutual information with known ground‑truth partitions, especially on networks that are intrinsically multiscale. The authors note that the method’s advantage stems from its explicit modeling of flow dynamics rather than static edge density alone.

Limitations discussed include the reliance on a random‑walk model, which may not perfectly represent specific dynamical processes such as epidemic spread or targeted information diffusion; the current implementation assumes static networks, so extensions to temporally evolving graphs are left for future work; and the automatic determination of the number of levels is driven solely by description‑length minimization, which could be complemented by domain‑specific constraints.

Future research directions proposed are: (i) incorporating empirical flow data (e.g., passenger counts, traffic sensors) to refine the transition probabilities; (ii) developing online or incremental versions of the algorithm for streaming network data; (iii) creating visualization tools that exploit the hierarchical codebooks to produce multiscale network maps; and (iv) analytically linking hierarchical compression to functional properties such as robustness, controllability, and diffusion speed.

In summary, the hierarchical map equation provides a principled, scalable, and dynamics‑aware approach to uncovering multilevel community structure in large integrated systems. By treating the compression of a random‑walk description as the objective, the method simultaneously discovers the optimal hierarchy, the appropriate number of levels, and the modular partition at each level, offering a powerful new lens for interpreting the organization of social, biological, and infrastructural networks.