.png){kind=icon}
.png){kind=icon}
Return
Optimization of model parameters and hyperparameters in deep learning models for spatial interaction prediction
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2024.126160.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The development of deep learning (DL) provides new methods for spatial interaction prediction. However, DL model training faces the challenges of gradient vanishing/exploding and falling into local optima. Parameters tuning is crucial to improve model performance. Currently, research on spatial interaction prediction pays little attention to the parameter tuning of neural network models. This research optimizes two types of constructed spatial interaction prediction models from model parameters and hyperparameters, respectively. They are spatial interaction prediction model based on BP neural network (BP-SIP) and spatial interaction prediction model based on spatio-temporal graph convolutional network (STGCN-SIP). We propose a new algorithm named Chaotic Differential Evolutionary Grey Wolf Algorithm (CDG) for the optimal solution, which improves the Differential Evolution Grey Wolf Optimization (DEGWO) algorithm. For the BP-SIP model, the research optimizes the initial weights and thresholds using CDG. For the STGCN-SIP model, the research employs Bayesian Optimisation (BO) and Adaptive Genetic Algorithm (AGA) to find the optimal hyperparameters combination, respectively, and compares their results. Experiments based on the Tencent migration data show that the proposed methods effectively improve the accuracy of models. The CDG algorithm reduces the MSE and RMSE of the BP-SIP model by 22.6% and 20% respectively, and increases R2 by 12.0%. The BO and AGA respectively reduce the MSE of the STGCN-SIP model by 26.2% and 24.2%, with R2 increasing by 20.6% and 15.0%. The results show the necessity and sufficiency of metaheuristics parameter optimization.
Keywords:
Deep learning model
Parameter optimization
Differential Evolution Grey Wolf Optimization
Bayesian Optimization
Adaptive Genetic Algorithm
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W