HYBRID DIAMOND CUT CLASSIFICATION AND PRICE PREDICTION USING FFNN BASED RF-CART FRAMEWORK

Abstract

Diamonds are a highly valued commodity, where accurate quality classification and price prediction are critical for both consumers and retailers. Market analytics indicate that over 80% of diamond transactions occur via digital platforms, yet more than 40% of price estimations lack consistency due to subjective grading and market volatility. Existing systems often struggle to map categorical quality indicators, such as cut, to price, hindered by class imbalance and non-linear data complexities. The dataset used in this study includes attributes such as carat, cut, color, clarity, depth, table, x, y, z, and price, with cut as the classification target. Traditional models show limited performance due to skewed class distributions and weak feature interactions. To address these challenges, an advanced pipeline was developed. Exploratory Data Analysis (EDA) was conducted to uncover patterns and correlations between quality attributes and price. Data preprocessing ensured normalization and encoding of categorical features, while Support Vector Machine–Synthetic Minority Oversampling Technique (SVM-SMOTE) was applied to balance the five cut classes: Fair, Good, Ideal, Premium, and Very Good. For predictive modelling, several hybrid Classification and Regression Tree (CART) models were tested, including Linear-CART, K-Nearest Neighbours–CART (KNN-CART), and a novel Feedforward Neural Network (FFNN) combined with Random Forest–CART (RF-CART). The proposed FFNN+RF-CART achieved 97.57% accuracy, 97.35% precision, 97.27% recall, and a 97.31% F1-score in cut classification. For price prediction, it delivered a Mean Absolute Error (MAE) of 0.1570, Mean Squared Error (MSE) of 0.1954, RMSE of 0.0004, and an R² of 0.9870. This integrated framework offers a robust, data-driven approach for real-time diamond evaluation, enhancing transparency and efficiency in the gemmological industry