Year: 2025
Type: article
Abstract: In regression forecasting problems based on large-scale and noisy datasets, there is often a need to choose between classical machine learning algorithms and modern neural network methods. Classical methods are simpler and more interpretable, while neural networks are better at handling heterogeneous and high-dimensional data, although they require more resources and more difficult fine-tuning. This paper presents a comparative analysis of the Random Forest (RF), XGBoosting, and Dense Neural Network (DNN) regression models for processing large tabular datasets. In particular, the IMDb dataset from the Kaggle platform was analyzed. Special attention was focused on studying the possibility of improving the performance of the prediction by combining RF and XGBoosting ensemble methods with DNN models. It was found that the RF model demonstrated acceptable predictive quality, namely, a coefficient of determination (R²) was 0.8640. The XGBoosting-based model showed a considerably better result, with an R² of 0.9245. The basic DNN model was characterized by the R² value of 0.8990. After optimizing the hyperparameters of the DNN model, the R² increased to 0.9179. A hybrid approach has been proposed as an additional way to improve the effectiveness of the DNN model. In particular, the distributions of features according to their impact on the prediction accuracy determined by the RF and XGBoosting methods were used as weighting coefficients for the DNN model feature vector. As a result, the most accurate forecast was obtained. The coefficients of determination R² were 0.9283 and 0.9302 for the RF-DNN and XGBoosting-DNN hybrid models, respectively. The obtained results can be used to develop predictive models based on heterogeneous and high-dimensional tabular data
Source: Artificial Intelligence
Institution: Lviv University
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: bronze
OpenAlex ID: W4411612849