A Purification-Enhanced Graph Neural Network Pipeline for Archaeological Provenance of Iberian Pottery

Determining the provenance of ancient ceramics remains a central challenge in Iberian archaeology due to the impurity of baseline reference datasets and the omission of spatial connectivity in traditional statistical models. To resolve these limitations, this study presents a two-phase decoupled computational pipeline designed to systematically isolate indigenous production from exotic trade items and reconstruct regional exchange networks across multiple historical Portuguese sites. Phase I introduces an unsupervised, dual-perspective parallel purification framework to establish high-confidence Local Reference Groups without introducing manual filtering bias. Rather than relying on a single metric, the pipeline deploys two independent and complementary models to cross-check anomalous data vectors. First, Hierarchical Density-Based Spatial Clustering (HDBSCAN) isolates outliers based on pure chemical element density distributions. In parallel, a Multilayer Perceptron Autoencoder (MLP-AE) projects the compositional vectors onto a unit hypersphere via L2 normalisation. This network evaluates reconstruction errors to detect subtle techno-chemical deviations in underlying crafting recipes and workshop traditions. Intersecting these density-based and reconstruction-based boundaries algorithmically purges systemic noise, protecting the downstream inference engine from training corruption. In Phase II, a supervised Graph Attention Network (GAT) is trained exclusively on these purified local datasets to learn joint geographic-chemical representations. By transforming individual shards into spatial-chemical topological graphs where landscape realities act as explicit network constraints, the model overcomes the analytical limitations of traditional point-cloud classifiers that evaluate observations in a geographical vacuum. Finally, the trained topological model is deployed to predict the specific sourcing origins of the isolated exotic shards and re-examine marginal local classifications. By extracting the multi-head attention coefficients and graph readout matrices, this pipeline delivers quantitative, interpretable maps of potential trade corridors and socioeconomic interaction spheres for empirical archaeological verification.