An Improved Hybrid Geometric Similarity-Based Retrieval Model for Case-Based Reasoning Systems

Abstract

This research proposed an improved model for case retrieval in Case-Based Reasoning (CBR) systems by introducing a hybrid geometric similarity measure. Traditional similarity measures often fail to jointly capture magnitude, orientation, and feature significance, hence leading to suboptimal case retrieval, particularly in domains requiring high precision. The proposed hybrid geometric similarity measure (NHSim) normalizes Euclidean distance and Cosine similarity before combining with Jaccard similarity by using a simple average approach. The resulting hybrid similarity formula has scores in the range of [0, 1] and was integrated with the kNearest Neighbors (k-NN) algorithm to evaluate case similarity retrieval. The model provides consistent, interpretable, scalable, and computationally lightweight similarity values. This approach leverages the strengths of all three measures while minimizing their individual limitations. The model was evaluated using a dataset of 45 catfish disease cases characterized by 33 binary symptoms, with a case base (60%) and query set (40%). Python was employed for the implementation, while evaluation metrics such as accuracy, precision, recall, F1-score, and AUC were used to measure the model's performance. Using a k value of 4 and threshold (τ) of 6.5, experimental results showed that the proposed hybrid model consistently outperformed individual similarity metrics, achieving 85% accuracy, 77% precision, 85% recall, and 80% F1-score. The AUC was 92%, highlighting the model’s robustness and reliability. The proposed model demonstrated significant improvements in retrieving relevant cases, particularly for binary-featured domains such as veterinary diagnostics. The findings suggest that hybrid similarity measures offer better accuracy than standalone metrics.