The eNOVATION Recommendation Engine is an ontology-driven, explainable decision-support system that assists CBRN professionals in identifying the most suitable Training Centres (TCs) and technologies for a given operational need. Operating directly over the eNOVATION ontology developed in T4.1, it delivers ranked recommendations accompanied by traceable semantic explanations, making every suggestion transparent and interpretable by domain experts.

The system exploits the TBox/ABox architecture of the eNOVATION ontology. The TBox (schema) drives automatic criterion discovery and meta-path enumeration, while the ABox (instance data) is used to validate path support and compute per-instance scores.

Bidirectional adjacency lists with subclass-aware matching and branch re-entry detection enable efficient graph traversal without redundant loops.

Key components include:

• Symbolic meta-path scoring: expert-defined semantic paths over the knowledge graph with multi-criteria aggregation, providing directly interpretable explanations for each recommendation.
• Knowledge Graph Embeddings: TransE and RotatE models trained on the eNOVATION knowledge graph for link prediction-based technology recommendation.
• Automatic criterion extraction: TBox-guided enumeration and ABox-validation of scoring criteria, enabling generalised recommendations without manual path specification for each new query type.
• Multi-level explainability: each recommendation exposes overall score, per-criterion group scores, natural-language justifications and instance-level graph paths normalised for readability.

Preliminary results showcase that the symbolic and embedding-based approaches are complementary: embeddings provide broader predictive coverage, while meta-path methods deliver superior interpretability. The generalised method achieves comparable ranking quality to the expert-designed approach without requiring per-query schema knowledge.

Key functionalities

• Symbolic meta-path scoring with multi-criteria aggregation
• Knowledge graph embeddings (TransE and RotatE) for link prediction
• Automatic criterion extraction (TBox-guided and ABox-validated)
• Multi-level explainability with scores and natural-language justifications

Available for Joint Activities

• VESTA – Belgium-Ranst (04/2027)

Contact

ICCS · Vagelis Karakolis
vkarakolis@epu.ntua.gr