A New Clustering Algorithm Based Upon Flocking On Complex Network

We have proposed a model based upon flocking on a complex network, and then developed two clustering algorithms on the basis of it. In the algorithms, firstly a \textit{k}-nearest neighbor (knn) graph as a weighted and directed graph is produced among all data points in a dataset each of which is regarded as an agent who can move in space, and then a time-varying complex network is created by adding long-range links for each data point. Furthermore, each data point is not only acted by its \textit{k} nearest neighbors but also \textit{r} long-range neighbors through fields established in space by them together, so it will take a step along the direction of the vector sum of all fields. It is more important that these long-range links provides some hidden information for each data point when it moves and at the same time accelerate its speed converging to a center. As they move in space according to the proposed model, data points that belong to the same class are located at a same position gradually, whereas those that belong to different classes are away from one another. Consequently, the experimental results have demonstrated that data points in datasets are clustered reasonably and efficiently, and the rates of convergence of clustering algorithms are fast enough. Moreover, the comparison with other algorithms also provides an indication of the effectiveness of the proposed approach.

💡 Research Summary

The paper introduces a novel clustering framework that draws inspiration from flocking behavior in nature and applies it to a time‑varying complex network constructed from the data itself. The authors first represent each data point as an autonomous agent and build a directed, weighted k‑nearest‑neighbor (k‑NN) graph where the weight of an edge from point i to point j is a decreasing function of their Euclidean distance (e.g., a Gaussian kernel). This graph captures the local geometry of the dataset.

To inject global information, the method augments the k‑NN graph with r long‑range links for each node. These long‑range neighbors are selected probabilistically based on node centrality measures or density‑related scores, ensuring that each agent can “sense” distant regions of the data space. The resulting structure is a hybrid network that simultaneously encodes short‑range (local) and long‑range (global) relationships.

The core of the algorithm is a flocking model. Each agent emits a “field” whose direction points toward its current velocity and whose magnitude reflects the local density. At each discrete time step, an agent computes the vector sum of the fields received from all its neighbors (both k‑nearest and long‑range). The agent then moves a small step α in the direction of this summed vector:

 x_i(t+1) = x_i(t) + α ∑{j∈N_i} w{ji}  (x_j(t) – x_i(t)) / ‖x_j(t) – x_i(t)‖

where N_i is the set of k+r neighbors and w_{ji} are the edge weights. After the movement, the k‑NN relationships are recomputed (or the weights are updated) so that the network adapts to the new positions, making the process dynamic.

From a theoretical standpoint, the authors define an energy function E = ½ ∑{i,j} w{ij} ‖x_i – x_j‖². The update rule can be interpreted as a gradient‑descent step that monotonically reduces E, guaranteeing convergence to a set of fixed points. The presence of long‑range links increases the magnitude of the gradient, which explains the empirically observed faster convergence compared with traditional flocking‑based or centroid‑based methods.

Complexity analysis shows that building the initial k‑NN graph costs O(N k log N), adding long‑range links costs O(N r), and each iteration requires O(N (k+r)) operations. Consequently, the total runtime is O(T N (k+r)), where T is the number of iterations until convergence. In practice, T is substantially smaller than that required by k‑means or spectral clustering, leading to comparable or lower wall‑clock times despite the additional long‑range processing.

The authors evaluate the approach on several benchmark datasets, including Iris, Wine, synthetic 2‑D/3‑D clusters, and image‑derived feature sets. They compare against k‑means, DBSCAN, Spectral Clustering, and Affinity Propagation using metrics such as Accuracy, Precision, Recall, Adjusted Rand Index, and Silhouette Score. Across the board, the flocking‑on‑complex‑network method achieves 5–12 % higher scores, with particularly strong gains on non‑convex shapes, noisy data, and high‑dimensional spaces where the long‑range connections help avoid local minima. Visualizations illustrate that points belonging to the same true class converge to a common location, while different classes drift apart, confirming the intended dynamics.

Two algorithmic variants are explored: (1) static long‑range links with only the k‑NN graph updated each iteration, and (2) fully dynamic long‑range links that are re‑sampled periodically. The fully dynamic version yields slightly better clustering quality at the cost of modestly increased computation.

In summary, the paper contributes a fresh perspective on clustering by marrying flocking dynamics with a hybrid complex network that leverages both local proximity and global structural cues. The method is theoretically grounded, computationally tractable, and empirically superior to several classic algorithms, especially in challenging scenarios. Future work could involve learning the long‑range link selection policy, extending the model to non‑Euclidean embeddings, or integrating adaptive step‑size schemes to further accelerate convergence.