Abstract
In this paper, we introduce the use of time-resolved variables to represent the evolution of categorical variables through time. Traditionally, in data science, the presence or absence of categorical variables are treated as binary (1 or 0) using One-Hot Encoding (OHE) and subsequently aggregated with compression techniques such as Principal Component Analysis (PCA). We annotated time-resolved variables on the Seshat dataset as sequences of categorical features and compared them to the same features treated with OHE and PCA compression. Our findings indicate that time-resolved variables are superior predictors for the evolution of social scale and social hierarchy, though they perform slightly less effectively for government specialization. We discuss the advantages and limitations of utilizing time-resolved variables for the computational analysis of cultural evolution.
Introduction: The Challenge of Quantifying History
The field of Cliodynamics—the quantitative study of history—seeks to identify the underlying drivers of social complexity by analyzing massive historical datasets. A primary resource in this endeavor is the Seshat Databank, which contains data on 186 polities across a 12,000-year window. However, historical data presents unique challenges for standard machine learning workflows. Most historical traits are categorical: a society either possesses a specific technology, such as "iron weapons," or it does not.
Standard practice involves converting these categories into binary flags via One-Hot Encoding. This approach suffers from two major drawbacks:
- High Sparsity: Because most societies lack most technologies for the majority of history, the resulting datasets are dominated by zeros, making them difficult for models to process efficiently.
- Information Loss in Compression: To manage thousands of binary variables, researchers often employ PCA. However, PCA is a "black box" technique that can obscure the nuances of historical progression, particularly when the underlying dimensions are low.
This paper proposes a novel feature engineering approach: the creation of "Chronos," a dataset that replaces sparse binary data with "time-resolved" ordinal scales.
Methodology: Time-Resolved Variables
The authors introduce time-resolved variables as a way to capture the cumulative nature of cultural evolution. Instead of treating related technologies as independent binary flags, they are arranged into a single numerical scale based on the chronological order of their first appearance in the global historical record.
For example, the Military Technology (militech) scale is constructed as follows:
- Level 1: Stone, wood, or clay tools.
- Level 2: Copper (approx. 5000 BC).
- Level 3: Bronze (approx. 3000 BC).
- Level 4: Iron (approx. 1400 BC).
- Level 5: Steel (approx. 400 BC).
- Level 6: Gunpowder (approx. 800 AD).
This logic was applied to ten macro-categories, including agriculture, religion, and information management. By mapping these traits onto developmental stages, the Chronos approach preserves the "evolutionary" sequence of history—knowing a society has reached Level 3 implies they have already mastered the technologies of Levels 1 and 2.
Comparative Analysis of Predictive Power
The effectiveness of the Chronos dataset was tested by its ability to predict three primary targets of social complexity: Social Scale (SCALE), Hierarchical Complexity (HIER), and Specialization of Governance (GOV).
As shown in the results above, the Chronos dataset outperformed the PCA baseline in linear regression tasks for both SCALE ($R^2$ of 0.744 vs 0.719) and HIER ($R^2$ of 0.829 vs 0.796). While the PCA-based model performed slightly better on "Specialization of Governance" (GOV), this is likely due to a "home-field advantage": the variables used to calculate GOV in the original Seshat data were themselves derived from OHE, favoring the PCA approach.
The Role of Dimensionality in Feature Compression
A critical insight from this research is that the predictive power of PCA-compressed variables is heavily tied to their initial dimensionality.
When dealing with high-dimensional data, such as warfare tactics (38 features), PCA retains strong predictive power. However, for low-dimensional categories like military technology (5 features), PCA compression significantly degrades performance. The Chronos approach mitigates this by encoding the cumulative information of these features into a single, semantically transparent variable, which maintains high correlation strengths with social complexity.
Discussion and Conclusion
The use of time-resolved variables offers significant advantages for historical analysis. Unlike abstract PCA components, Chronos variables are semantically transparent; a specific level on a scale corresponds to a concrete set of historical capabilities, making models far more interpretable for historians.
However, the method is not a universal solution. It struggles with traits that do not follow a linear evolutionary path. For instance, writing systems and political structures often exhibit non-linear patterns—cultures may adopt a system and maintain it for millennia or oscillate between different forms of governance. Future work will focus on expanding the Chronos dataset to include more recent history and incorporating textual data to better quantify the reliability of historical records.
Summary updated