Analysis of Criminal Patterns in Police Report Narratives using Spectral Clustering with K-means

Authors:
Publisher website
Google Scholar
Abstract
The high volume and heterogeneity of police report narratives in Brazil pose challenges for manual analysis and investigative prioritization. This work proposes an approach for identifying criminal patterns using clustering techniques applied to unstructured textual data. The methodology integrates Spectral Clustering with K-means, leveraging MPNet embeddings for vector representation and UMAP for dimensionality reduction. The resulting six clusters revealed thematic coherence, highlighting patterns such as bank fraud, judicial scams, social media crimes, and account hacking. Comparative experiments with Agglomerative Clustering were conducted using different linkage strategies, with Spectral Clustering achieving the highest silhouette score (0.38), indicating better-defined groups. A manual qualitative analysis of samples from each cluster supported the thematic distinctions. The study demonstrates that automatic clustering can contribute to investigative triage, offering relevant insights for public security applications.