01 Project
GRAPH MACHINE LEARNING · KNOWLEDGE REPRESENTATION
Graph Intelligence
Ontology-Aware Classification of Archaeological Ceramics
“Can graph neural networks leverage structured domain knowledge to classify archaeological artifacts with incomplete attributes?”
Developed an ontology-aware graph learning pipeline that represents ceramic objects, material properties, and functional relationships as a heterogeneous graph, comparing relational graph neural networks against a conventional tabular MLP baseline.
The Problem & Context
Bridging incomplete physical artifact records with structured domain knowledge representations.
Classifying incomplete archaeological records
Archaeological ceramics provide valuable information about historical production techniques, material composition, and cultural practices. Their classification relies on specialized domain knowledge and observations of physical characteristics.
However, archaeological records are often incomplete. Depending on an object's preservation and documentation, attributes such as fabrication techniques, surface treatments, or intended functions may be missing.
Traditional classification approaches generally represent each artifact as an independent feature vector, without explicitly modeling the relationships between objects and the broader archaeological ontology.
From catalog records to structured knowledge
The project used a collection of more than 11,000 archaeological ceramic records alongside an existing domain ontology.
The available information included material properties, manufacturing characteristics, functional descriptions, and a hierarchical taxonomy of technical categories.
Rather than relying exclusively on a flat representation, the project explored how this domain knowledge could be translated into a heterogeneous graph.
Representing Archaeological Knowledge as a Graph
Translating tabular observations and ontological taxonomies into an interconnected heterogeneous graph.
Ceramic Object
Individual physical artifacts with partial or noisy documentation.
has_feature →Observed Feature
Typed properties: fabric, forming method, firing, and surface coating.
subClassOf →Functional Entity
Hierarchical functional taxonomy connecting related concepts.
categorized_as →Technical Category
Target root technical classification labels for model prediction.
• Target LabelThe proposed representation combines four types of entities: ceramic objects, features, functions, and technical categories. Ceramics are connected to observed characteristics through typed relationships.
Node representations are initialized using semantic embeddings generated from archaeological descriptions and domain terminology, providing informative dense inputs.
Target Leakage Prevention: For the classification task, target-category associations are strictly isolated as supervision labels rather than ordinary graph connections available during relational message passing.
Learning from Relationships and Attributes
Two complementary machine learning approaches were implemented and evaluated under identical conditions.
Relational Graph Neural Network
A Relational Graph Convolutional Network (R-GCN) processes the heterogeneous graph using relation-specific message passing.
Aggregates neighborhood features according to distinct edge relations (fabric, firing, coating).
Combines initial semantic embeddings with multi-hop connectivity to fill missing attributes.
Learned ceramic embeddings are passed to a dense classification head predicting technical categories.
Tabular MLP Baseline
A conventional Multi-Layer Perceptron trained on encoded ceramic attributes, features, and functional info.
Encodes physical observations and domain characteristics as independent feature vectors.
Processes records without graph connectivity or topological neighborhood propagation.
Provides a direct benchmark to evaluate whether graph complexity produces genuine performance gains.
Measuring Classification Performance
Evaluating multi-sampling configurations, class distribution sensitivity, and metric trade-offs.
| Evaluation Metric | R-GCN Pipeline (Graph) | Tabular MLP (Baseline) | Relative Improvement |
|---|---|---|---|
| Classification Accuracy | 98.33% | 91.20% | +7.13% |
| Macro-Averaged F1 | 0.978 | 0.894 | +0.084 |
| Macro-Averaged Precision | 0.981 | 0.902 | +0.079 |
| Macro-Averaged Recall | 0.976 | 0.887 | +0.089 |
Technical Challenges & Research Takeaways
Key obstacles encountered in heterogeneous graph modeling and the methodological takeaways.
Incomplete & Heterogeneous Information
Input records contained varied attributes with different degrees of completeness. Designing useful representations required consolidating data from multiple source tables while preserving relevant archaeological characteristics.
Translating Domain Ontology to Graph
The domain ontology included hierarchical concepts and multiple relationship types. Converting this into a machine-learning graph required consistent entity mappings, typed relation definitions, and unified node embeddings.
Preventing Target Information Leakage
The classification task required rigorous separation between relationships available to the model and the technical-category associations used as prediction targets. The pipeline isolated message-passing edges from supervision links.
When Domain Knowledge Becomes Useful
Representations tailored to problems: This project reinforced the importance of choosing representations according to the underlying problem rather than the complexity of the model.
Justifying architectural complexity: Graph-based learning provides a natural way to encode shared properties and structured relationships, but its added complexity must be justified through comparisons with simpler baselines.
Methodological rigor: Highlighted the importance of reliable experimental protocols, appropriate class-sensitive evaluation metrics, and careful handling of hierarchical taxonomies.
Connecting ML with domain expertise: Demonstrated how modern machine learning can be deeply coupled with specialized domain knowledge to address problems involving heterogeneous and incomplete observations.
Source Code & Implementation
Explore the Spiridon research codebase, experimental setup, and technologies.
Spiridon Repository
The Spiridon repository contains the complete data-preparation pipeline, ontology-based graph construction, R-GCN classifier, tabular MLP baseline, and experimental configuration.