Graph Max Shift: A Hill-Climbing Method for Graph Clustering

We present a method for graph clustering that is analogous to gradient ascent methods previously proposed for clustering points in space. The algorithm, which can be viewed as a max-degree hill-climbing procedure on the graph, iteratively moves each node to a neighboring node of highest degree. We show that, when applied to a random geometric graph whose nodes correspond to data drawn i.i.d. from a density with Morse regularity, the method is asymptotically consistent. Here, consistency is in the sense of Fukunaga and Hostetler, meaning, with respect to the partition of the support of the density defined by the basins of attraction of the density gradient flow.

💡 Research Summary

The paper introduces a novel graph clustering algorithm called Graph Max Shift, which operates by iteratively moving each vertex to a neighboring vertex with the highest degree. This simple hill‑climbing rule is motivated by the continuous‑space mean‑shift and max‑shift procedures that climb a density surface toward its modes. The authors bridge the gap between the continuous and discrete settings by showing that, when the input graph is a random geometric graph generated from points sampled i.i.d. from a smooth density f on ℝ^d, the degree of a vertex is proportional to a flat‑kernel density estimator evaluated at the latent point. Consequently, the path taken by Graph Max Shift coincides exactly with the path taken by the classical Max‑Shift algorithm applied to the latent point set, and the terminal vertices of these paths correspond to the modes of f.

The theoretical contribution is a rigorous proof of statistical consistency: under standard regularity assumptions (f is a Morse function, twice differentiable with non‑singular Hessian at critical points, and decays to zero at infinity) and a connectivity radius ε_n that shrinks to zero slowly enough (specifically, n ε_n^d / log n → ∞), the partition produced by Graph Max Shift converges in probability to the partition defined by the basins of attraction of the gradient‑ascent flow of f, as originally defined by Fukunaga and Hostetler. The authors quantify convergence using the Rand index (or equivalently one minus the Rand distance) between the algorithm’s output partition and the population partition, showing that this distance tends to zero as the sample size grows.

The algorithm is described in detail: compute the degree q_i for each vertex i, then for each starting vertex i_0 repeatedly set i_{t+1}=arg max_{j∈N(i_t)} q_j, breaking ties deterministically (e.g., by choosing the larger index). The process stops when no neighbor has a higher degree; all vertices whose paths terminate at the same “terminal node” are grouped together. An optional post‑processing step merges clusters whose terminal nodes are within τ hops of each other; τ is the sole hyper‑parameter.

The paper situates Graph Max Shift among related work. It clarifies that traditional graph partitioning methods (Kernighan‑Lin, spectral cuts, etc.) aim to minimize global cut costs, whereas Graph Max Shift exploits local degree information to approximate a density landscape. It also connects to Max‑Shift (the flat‑kernel version of mean‑shift) by showing exact equivalence on geometric graphs, and contrasts with Max‑Slope Shift (Könz et al.), which uses a degree‑over‑distance ratio and converges to density‑gradient inflection points rather than true modes. Moreover, the authors demonstrate that Graph Max Shift can be expressed as a special case of Morse clustering, defined via node and edge pre‑orders, thereby linking to the axiomatic clustering literature (Kleinberg’s impossibility theorem and subsequent work by Strazzeri & Sánchez‑García).

Empirical validation is performed on synthetic datasets of varying dimension (d = 2, 3, 5) and mixture densities (multiple Gaussian components, mixed Bernoulli, etc.). Random geometric graphs are constructed with radii chosen to satisfy the theoretical scaling regime. The authors compare Graph Max Shift against spectral clustering, k‑means on graph embeddings, and Max‑Slope Shift, measuring performance with the Rand index. Graph Max Shift consistently attains the highest Rand scores, confirming its ability to recover the true modal partition. The τ‑merging step is shown to effectively control over‑segmentation, aligning the number of discovered clusters with the true number of modes.

Complexity analysis reveals that degree computation costs O(|E|) and each vertex’s hill‑climbing path requires a single scan of its adjacency list, yielding an overall time complexity of O(n · \bar q), where \bar q is the average degree. For sparse geometric graphs (ε small enough that \bar q remains O(1)), the algorithm runs essentially in linear time. Memory usage is O(|E|) using adjacency lists. The authors provide a reference implementation on GitHub, facilitating reproducibility.

In conclusion, Graph Max Shift offers a conceptually simple, computationally efficient, and theoretically sound method for clustering graphs when the underlying data are generated from a smooth density. It achieves asymptotic recovery of the gradient‑ascent basins without requiring any embedding or explicit density estimation. Future directions suggested include extensions to directed or weighted graphs, online versions for dynamic networks, incorporation of alternative local vertex scores (e.g., clustering coefficient, PageRank) into a multi‑criteria hill‑climbing scheme, and scalable distributed implementations for massive graphs.