"Halal or Not: Knowledge Graph Completion for Predicting Cultural Appropriateness of Daily Products"

Knowledge graphs transform scattered ingredient data into meaningful halal predictions.

HaCKG ( Halal Cosmetic Recommendation Framework), proposed in this paper, leverages knowledge graphs to predict halal status of cosmetics by learning complex relationships between products and ingredients through graph attention networks.

-----

https://arxiv.org/abs/2501.05768v1

🎯 Original Problem:

→ Image-based and text-based methods struggle with scientific ingredient names and complex formulations.

-----

🛠️ Solution in this Paper:

→ HaCKG constructs a cosmetic knowledge graph representing relationships between products, ingredients, and their properties.

→ A pre-trained relational Graph Attention Network (r-GAT) with residual connections learns structural relationships in the knowledge graph.

→ The model fuses numerical and categorical attributes through a gate function for unified feature representation.

→ A two-phase training approach uses self-supervised pre-training followed by fine-tuning for halal prediction.

-----

💡 Key Insights:

→ Knowledge graphs effectively capture complex relationships between cosmetics and ingredients

→ Pre-training on graph structure improves model robustness

→ Residual connections prevent over-smoothing in deep graph networks

-----

📊 Results:

→ HaCKG achieves 96.57% accuracy in halal prediction

→ Outperforms baseline models by 9-33% across all metrics

→ Maintains performance with just 2-3 graph layers