Structural Pattern Mining in Inka Khipus: Unsupervised Clustering, Provenance Classification, and Computational Validation of the Santa Valley

Jul 16, 2026

Upload by Anda

Khipus—knotted cord devices of the Inka Empire—remain largely undeciphered, prompting a reproducible machine-learning pipeline on the Open Khipu Repository (OKR). The study engineers 27 structural features per khipu and applies UMAP and HDBSCAN to recover three structurally distinct groups (silhouette = 0.769). It further trains a gradient-boosting provenance classifier (F1 = 0.86) for Inka Late Horizon imperial style and uses SHAP interpretability to highlight cord twist direction as the dominant discriminator. The work also computationally validates the Santa Valley moiety/recto-verso structure and reports that n-gram knot-type sequence order adds no extra provenance signal beyond aggregate features.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages10

Reading Time25 min

Introduction

Related Work

Dataset and Feature Engineering

The Open Khipu Repository

Feature Engineering Pipeline

arXiv :2607 .00185v1 [ cs .CL] 30 Jun 2026 Structural Pattern Mining in Inka Khipus: Unsupervised Clustering, Provenance Classification, and a Computational Validation of the Santa Valley Match Maria Contreras 1 1 Universidad Peruana de Ciencias Aplicadas (UPC), Lima, Peru, [mariacontrerasdsg@outlook. com](mariacontrerasdsg@outlook. com) July 2, 2026 Abstract Khipus—knotted cord devices—were the primary recording medium of the Inka Empire (c. 1400–1532 CE), yet their system remains undeciphered. We present a reproducible machinelearning pipeline applied to the Open Khipu Repository (OKR), a public database of 619 khipus comprising 54,403 cords and 110,677 knots. We engineer 27 structural features per khipu and apply (i) unsupervised clustering via UMAP and HDBSCAN, recovering three structurally distinct groups (silhouette = 0 .769); (ii) supervised provenance classification via gradient boosting, reaching F1 = 0 .86 for the Inka Late Horizon imperial style; and (iii) SHAP-based interpretability, which identifies cord twist direction as the dominant structural discriminator of imperial khipus. We further report two findings of methodological interest. First, one cluster is dominated not by a geographic region but by nineteenth-century European museum collections, indicating that colonial acquisition and recording practices are structurally encoded in the corpus. Second, we provide an independent computational verification of the recto/verso (moiety) structure of the six Santa Valley khipus reported by Medrano and Urton (2018), reproducing both the aggregate attachment ratio and the identification of the single mixed specimen—using only the public OKR database, without physical access to the objects. We additionally report a negative result: knot-type sequence order, encoded as n-grams, adds no provenance signal beyond aggregate features. All code and data are openly available. Keywords: computational archaeology, khipu, quipu, unsupervised learning, UMAP, HDBSCAN, gradient boosting, SHAP, digital humanities, Andean studies 1 Introduction The Inka Empire was the largest polity in the pre-Columbian Americas, administering millions of people across the Andes without a writing system in the conventional sense. In its place, the Inka state relied on khipus (Quechua for “knot”): assemblages of spun and plied cotton or camelidfiber cords, knotted to record numerical and possibly narrative information. Spanish chroniclers described specialist record-keepers, the khipukamayuqs, who read these devices as administrative documents. Roughly 1,000 khipus survive in museum and private collections worldwide. While the baseten numerical convention of many khipus is well understood (Ascher and Ascher, 1997 ; Urton, 2003), whether and how they encode language, names, or narratives remains one of the principal open problems in Andean studies. Most quantitative work to date has focused either on individual specimens or on matching numerical sums across related khipus (Urton and Brezine, 2005) . The structural dimensionality of the full corpus—what construction conventions distinguish khipus across regions and periods—has received less systematic computational attention. This paper applies modern machine-learning methods to the Open Khipu Repository (OKR), the most comprehensive public khipu database. Our contributions are: 1. A reproducible feature-engineering pipeline that converts the OKR’s relational structure into a 27-dimensional vector per khipu (Section 3) . 2. An unsupervised analysis (UMAP + HDBSCAN) revealing three well-separated structural clusters, one of which reflects colonial collection bias rather than geography (Section 4) . 3. A supervised provenance classifier with interpretability analysis, identifying cord twist direction as the signature of Inka imperial manufacture (Section 5) . 4. An independent computational verification of the Santa Valley moiety structure reported by Medrano and Urton

FAQs

What dataset and scale are used in the study?

The pipeline is applied to the Open Khipu Repository (OKR), containing 619 khipus, 54,403 cords, and 110,677 knots aggregated from multiple prior efforts.

How are structural patterns discovered and how many clusters are recovered?

The method engineers 27 structural features per khipu and runs unsupervised clustering using UMAP and HDBSCAN, recovering three structurally distinct groups with silhouette = 0.769.

What cues determine imperial provenance in the supervised model?

A gradient-boosting classifier achieves F1 = 0.86 for Inka Late Horizon imperial style, and SHAP-based interpretability identifies cord twist direction as the dominant structural discriminator.

Jul 16, 2026

Upload by Anda