In materials science, the reliability of predictive modeling heavily depends on data completeness. Unfortunately, datasets often have missing values due to experimental limitations or measurement errors, compromising model accuracy. This study introduces an optimized K-Nearest Neighbors (KNN) imputation technique combined with Deep Neural Network (DNN) modeling to enhance predictions of material properties.
Improved Imputation Techniques Traditional imputation methods like mean imputation and Multiple Imputation by Chained Equations (MICE) can inadequately handle complex datasets. The proposed Enhanced KNN approach optimizes the number of neighbors and employs refined distance metrics to ensure minimal bias during imputation. This method demonstrates superior performance with an R² score of 0.973, outperforming mean imputation, MICE, and even the standard KNN.
Integration with Deep Learning After imputation, the completed dataset is used to train a DNN, which accurately predicts material properties such as density. The model's architecture, with multiple hidden layers and tailored activation functions, ensures high precision. Evaluated using Mean Squared Error (MSE) and Mean Absolute Percentage Error (MAPE), the DNN shows significant predictive capabilities, underscoring the potential of integrating advanced imputation methods with deep learning.
Key Takeaways The study highlights that using enhanced imputation techniques preserves essential data characteristics and improves the performance of machine learning models. This advancement is crucial for accurate material property predictions, fostering further innovation in materials research and applications.
🔗 Full Text: https://www.igminresearch.com/articles/html/igmin197
🔗 DOI Link: https://dx.doi.org/10.61927/igmin197
Comments
Post a Comment