Supplier Selection Recommendation System: A Comparative Study of K-Nearest Neighbors, Random Forest, and Dual-Encoder Models

Supplier selection is an important supply chain activity traditionally reliant on subjective, simple scoring methods and human judgment. This thesis frames selection as a recommendation problem, similar to how online platforms suggest products. Context features were utilized to identify optimal supplier matches. An offline evaluation compared K-Nearest Neighbors (K-NN), Random Forest, and Dual Encoder models across temporal and random data splits. Results indicate that K-NN achieved the highest ranking accuracy, proving that contract similarity is a powerful proxy for relevance in sparse datasets. Conversely, the Dual Encoder showed lower accuracy but wider coverage, highlighting a greater ability to generalize well. The Random Forest performed poorly due to sparse feature sets. The findings show that model performance is limited by unseen suppliers and reveal a key trade-off between ranking precision and supplier exploration. Ultimately, aligning model architecture with specific procurement objectives is essential as an effective decision support system.