Partitioning Israeli Municipalities into Politically Homogeneous Cantons: A Constrained Spatial Clustering Approach
Algorithmic analysis of 229 municipalities reveals five politically coherent regions with 0.905 silhouette score.
Researchers Adir Elmakais and Oren Glickman have published a novel computational study applying spatial clustering algorithms to Israel's political geography. Their paper, 'Partitioning Israeli Municipalities into Politically Homogeneous Cantons: A Constrained Spatial Clustering Approach,' analyzes municipality-level election results from five Knesset elections (2019-2022) to explore hypothetical divisions of the country into politically coherent regions. The methodology tests four clustering algorithms—Simulated Annealing, Agglomerative Clustering with contiguity constraints, Louvain Community Detection, and K-Means as baseline—across 264 experimental configurations involving different feature representations, distance metrics, and cluster counts.
Key findings reveal that Agglomerative clustering using BlocShares features with Euclidean distance achieved the highest clustering quality with a silhouette score of 0.905. Meanwhile, Non-negative Matrix Factorization (NMF) with Louvain community detection provided the best balance between political homogeneity and interpretable canton assignments. The research demonstrates remarkable temporal stability, with deterministic algorithms producing near-perfectly stable partitions (Adjusted Rand Index up to 1.0) across all five elections, suggesting Israel's political geography remains structurally consistent despite electoral volatility.
The resulting optimal partition (K=5) identifies five politically coherent regions: a center-leaning metropolitan core, a right-wing southern arc, a right-leaning northern mixed region, and two distinct Arab-majority cantons. These divisions closely reflect known political-demographic patterns in Israeli society. The study includes an interactive web application that allows users to explore different partition configurations, making this data-driven approach to political geography accessible to researchers, policymakers, and the public interested in understanding polarization through computational methods.
- Tested 264 configurations across 4 clustering algorithms (Simulated Annealing, Agglomerative, Louvain, K-Means) with election data from 229 municipalities
- Agglomerative clustering with BlocShares features achieved highest quality (silhouette score 0.905), while NMF with Louvain provided best balance
- Identified 5 politically coherent cantons that remain stable across 5 elections (2019-2022) with ARI up to 1.0
Why It Matters
Provides data-driven framework for understanding political polarization through computational methods, with applications for governance and social science research.