💻 computer science

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

This paper presents a reproducible machine-learning pipeline applied to the Open Khipu Repository that successfully clusters undeciphered Inka khipus into distinct structural groups, classifies their provenance with high accuracy, identifies cord twist direction as a key imperial marker, and computationally validates the moiety structure of the Santa Valley khipus without physical access to the artifacts.

Original authors: Maria Contreras

Published 2026-07-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Maria Contreras

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the Inka Empire as a massive, ancient organization that ran a country the size of modern-day Peru without using a single written word. Instead of paper and ink, they used khipus: intricate devices made of cotton or camelid hair cords, tied with knots to record numbers and stories. Think of them as the "hard drives" of the Inka bureaucracy.

For centuries, these devices have been a mystery. We know how to read the numbers (like counting sheep or taxes), but we don't know how they recorded names, stories, or languages.

This paper is like a team of digital detectives using modern computer tools to look at a giant public database of 619 khipus (containing over 110,000 knots) to see if they can find hidden patterns. They didn't try to "read" the language yet; instead, they asked: "Do these devices look different depending on where they were made or who made them?"

Here is what they found, explained simply:

1. The "Grouping" Game (Unsupervised Clustering)

The researchers used a computer algorithm to sort these 619 khipus into groups based on how they were built (how long the cords were, how many knots, and how the cords were twisted).

Imagine you have a huge pile of socks. You don't know who owns them, but you sort them by color and pattern. The computer found three distinct groups:

Group A (The "Imperial" Group): A small, very tight group of khipus that look almost identical. These are the "official" ones made by the Inka Empire's central government. They are like the standard-issue uniforms of an army.
Group B (The "Local" Group): The biggest group. These look like the everyday khipus used by local communities along the coast.
Group C (The "Museum" Group): This was a surprise. This group wasn't defined by geography (where they were made) but by where they ended up. These khipus were mostly collected by European museums in the 1800s. The computer realized that the way these specific museums recorded data (or failed to record data) left a "digital fingerprint" on the khipus. It's like realizing a group of socks all look different because they were all washed in the same strange machine in a museum basement, not because they came from a different town.

2. The "Imperial Signature" (Twist Direction)

The researchers then tried to teach a computer to guess where a khipu came from just by looking at its structure. They found one specific feature that was the "smoking gun" for the Imperial group: the direction the cords were twisted.

Think of twisting a rope. You can twist it clockwise (like a screw going in) or counter-clockwise.

The Imperial khipus were almost exclusively twisted in one specific direction (85% of the time).
The computer learned that if a khipu has this specific twist, it's almost certainly an official government document. It's like spotting a specific brand of thread that only the royal tailor used.

3. The "Santa Valley" Mystery Solved (Validation)

In 2018, other researchers claimed that six specific khipus from the Santa Valley were actually a record of taxes from 1670. They argued that the direction the cords were attached (some facing "front," some "back") encoded a social division (like "left side" vs. "right side" of a village).

The authors of this paper wanted to check if this was true without ever touching the physical objects. They went back to the public database and ran the numbers themselves.

The Result: They found the exact same pattern. Five of the six khipus were purely "front" or purely "back," and exactly one was a mix.
Why it matters: This proves that the pattern isn't a fluke. The structure of the cords really does match the social structure the researchers claimed. It's like verifying a recipe by cooking it yourself using only the written instructions, without ever seeing the original chef.

4. What They Didn't Find

The researchers also tried to see if the order of the knots mattered (like reading a sentence where the order of words changes the meaning). They tested this, but the computer said, "Nope, the order doesn't seem to tell us where the khipu came from." It's like trying to guess a person's hometown by the order of letters in their name, only to find that the order doesn't actually help.

The Big Picture

This paper doesn't "crack the code" of the Inka language. Instead, it built a map of the khipus.

It showed us that the Inka Empire had a very strict, standardized way of making these devices.
It warned us that our current database is biased because of how European museums collected them in the past.
It proved that we can verify complex historical theories using only digital data, without needing to hold the ancient artifacts in our hands.

In short, they used math to show that these ancient knotted strings have a hidden structure, and that structure holds the key to understanding how the Inka Empire was organized.

Technical Summary: Structural Pattern Mining in Inka Khipus

Problem Statement
The Inka Empire (c. 1400–1532 CE) utilized khipus—assemblages of knotted cords—as its primary recording medium. While the base-ten numerical conventions of these devices are understood, the encoding of language, names, or narratives remains undeciphered. Existing quantitative research has largely focused on individual specimens or specific numerical matches between related khipus. There is a lack of systematic computational attention regarding the structural dimensionality of the full corpus: specifically, what construction conventions distinguish khipus across different regions and periods, and how colonial acquisition practices may have biased the available data.

Methodology
The authors apply a reproducible machine-learning pipeline to the Open Khipu Repository (OKR), a public database containing 619 khipus, 54,403 cords, and 110,677 knots. The methodology proceeds in three stages:

Feature Engineering: The relational database structure is converted into a 27-dimensional feature vector for each khipu. These features aggregate cord- and knot-level attributes into four groups:
- Cord structure: Count, mean/standard deviation of length, maximum hierarchical level, and subsidiary-cord ratio.
- Twist: Counts and ratios of S- and Z-direction cords.
- Knots: Total count, mean turns, direction counts, and one-hot encoding of knot types (e.g., simple, long, figure-eight).
- Color: Number of unique colors, color entropy (measuring palette diversity), and a multicolor indicator.
  Missing values are imputed as zero, and features are standardized.
Unsupervised Clustering: The 27-dimensional vectors are reduced to two dimensions using UMAP (Uniform Manifold Approximation and Projection) and subsequently clustered using HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise). This approach requires no prior specification of the number of clusters.
Supervised Classification and Interpretability: A gradient-boosted tree classifier (XGBoost) is trained to predict geographic provenance from the structural features. The model is evaluated using five-fold stratified cross-validation, optimizing for the weighted F1 score. SHAP (SHapley Additive exPlanations) values are used to interpret feature importance, specifically for the "Inka Late Horizon" class.
Computational Validation: The study performs an independent verification of the recto/verso (moiety) structure of six Santa Valley khipus, previously analyzed by Medrano and Urton (2018), using only the public OKR data.

Key Results

Structural Clustering: The unsupervised analysis identifies three distinct structural clusters with a silhouette score of 0.769:
- Cluster 0: A small, highly cohesive group (17 khipus) corresponding to the Inka Late Horizon imperial style.
- Cluster 1: The majority corpus (442 khipus), predominantly from the Central Coast of Peru.
- Cluster 2: A group (160 khipus) dominated by European and North American museum collections (e.g., Museum für Völkerkunde, Berlin).
Colonial Bias Detection: Cluster 2 is not defined by geography but by acquisition history. Statistical analysis reveals a significant difference in twist-ratio features between Cluster 2 and Cluster 1 ( $p \approx 3.7 \times 10^{-276}$ ), largely driven by unrecorded twist data ("U") in specific colonial-era collections. This indicates that the OKR corpus contains structural biases reflecting 19th-century acquisition and recording practices.
Provenance Classification: The supervised classifier achieves a weighted F1 of 0.46 overall, limited by a small labeled sample (135 specimens) and class imbalance. However, the "Inka Late Horizon" class achieves an F1 of 0.86, confirming that imperial khipus follow highly standardized, machine-learnable construction conventions. SHAP analysis identifies cord twist direction as the dominant discriminator, with S-direction twist characterizing imperial khipus (85.3%). Other significant features include shorter/uniform cord lengths, higher knot counts, and lower color entropy.
Santa Valley Verification: The study independently reproduces the recto/verso structural pattern of the six Santa Valley khipus. The aggregate ratio of 49.0% recto to 51.0% verso matches the ~47%/53% ratio reported by Medrano and Urton. Crucially, the computational analysis identifies exactly one mixed specimen (KH0326) among five pure specimens, replicating the original study's finding without physical access to the objects.
Negative Result: Encoding knot-type sequence order as n-grams (bigrams) adds no provenance signal beyond aggregate features, yielding a negligible change in F1 ( $\Delta F1 \approx -0.006$ ).

Significance and Claims
The paper positions structural pattern mining not as a substitute for philological decipherment, but as a complementary tool for characterizing the corpus, generating testable hypotheses, and quantifying biases inherent in the data.

The primary contributions are:

A reproducible pipeline converting the OKR relational structure into a 27-dimensional structural vector.
The identification of three structural clusters, revealing that one major cluster reflects colonial collection bias rather than geographic origin.
The confirmation that Inka imperial khipus possess a distinct, standardized structural signature (dominated by cord twist direction) that is separable from regional styles.
The first independent computational verification of the Santa Valley moiety structure using only public database records.

The authors emphasize that these findings highlight the necessity of confronting collection biases in any future decipherment efforts and advocate for the use of indigenous terminology and open data practices to support descendant communities and Peruvian institutions.