Revolutionizing Urban Economic Competitiveness Analysis Using a CNN-based Approach

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Nov 9 2023

In a paper published in the journal PLoS ONE, researchers addressed the limitations of traditional urban economic competitiveness research methods, which often fail to capture the intricate relationships among urban features. They introduced a novel approach utilizing convolutional neural networks (CNN) and a complex urban feature dataset comprising 1008 features from 283 Chinese cities.

This approach not only provided a more comprehensive understanding of urban development but also addressed the issue of limited sample size by incorporating a data augmentation technique based on deep convolutional Generative Adversarial Networks (GANs), resulting in an improved model for analyzing regional development disparities.

China's Urban Economic Competitiveness

In China, fierce competition among cities for economic development has made understanding urban economic competitiveness crucial. Current research falls into two categories: causal factors affecting competitiveness and quantifiable indicators for ranking cities. However, these traditional methods assume linear relationships among factors and neglect complex interactions within urban systems. Deep learning, particularly CNN, offers a solution by modeling intricate urban features without prior assumptions. Despite its promise, CNNs require substantial training data.

Comprehensive Urban Feature System and Model Selection

The study focuses on constructing a comprehensive urban feature system to analyze economic competitiveness, considering economic, social, and environmental dimensions. This feature system includes 1067 indicators, categorized into three primary, 14 secondary, 42 tertiary, and 1008 specific indicators, sourced from official statistical data. Data for 283 Chinese prefecture-level cities over eight years (2012-2019) was collected, resulting in 2264 city samples. The urban comprehensive competitiveness index, which integrates various economic, social, and environmental factors, was the dependent variable.

Researchers chose CNNs as the suitable analytical model for handling this intricate dataset. CNNs excel at extracting local features and spatial correlations, making them well-suited for capturing the complex relationships among diverse urban indicators. The researchers employed GANs to augment the data, enhancing the model's performance by expanding the training dataset, improving generalization, and balancing data distribution.

GAN is a framework comprising two deep neural networks: a generator and a discriminator. The generator learns to create data resembling actual samples, while the discriminator aims to differentiate between accurate and generated data. They engage in a competitive training process, where the generator strives to create data indistinguishable from accurate data while the discriminator attempts to spot the difference. This adversarial training leads to the generation of high-quality synthetic data.

This study used the Deep Convolutional GAN (DCGAN) from various GAN variants. DCGANs employ CNN for both the generator and discriminator. The design of these networks is critical to ensure balanced training. The generator aims to produce realistic data, while the discriminator distinguishes natural from fake data. The architecture of the DCGAN involves deep convolutional layers, specific activation functions, and careful hyperparameter tuning to achieve this balance. It enables the model to generate data with realistic features, making the approach valuable for data augmentation and enhancing neural network performance.

Data Preprocessing, Evaluation, Results, and Generalization

Data Preprocessing and Analysis: Researchers discussed the data preprocessing and analysis methods used in the study by detailing how they standardized and normalized the dataset, which consisted of a wide range of indicators with various units of measurement. By applying normalization and standardization techniques, they transformed the data into a standard scale, eliminating the influence of different measurement units on the analysis. Additionally, the authors described the process of dividing continuous variables into discrete categories to improve model accuracy and efficiency. The researchers also address the distribution of sample labels and the class imbalances within the data.

Evaluation Metrics and Experimental Setup: The study employs several evaluation metrics to assess the performance of the CNN and the impact of data augmentation using DCGAN. Classification accuracy of the validation and test sets is the primary evaluation index for the CNN performance. The authors also discuss the importance of the cross-entropy loss function and convergence speed. The experimental setup, including the hardware and software used for model development, is provided for transparency.

Results and Robustness Analysis: The analysis focuses on presenting the study's results, both for the original dataset classification using CNN and the data augmentation using DCGAN. The authors discussed the improvements in classification accuracy achieved by augmenting the dataset and highlight the quality assessment of the generated data. They also analyzed the robustness of their approach by comparing it with other neural network architectures, such as Visual Geometry Group (VGG), Vision Transformer (ViT), and Long Short-Term Memory (LSTM), to validate the proposed CNN model's performance. This article provides a comprehensive overview of the study's findings and implications.

Generalization and Data Augmentation Validation: This study demonstrates the adaptability of the CNN-Custom model by applying it to county-level economic data in China, achieving promising results in the 5-class classification. However, challenges arise when increasing the classes to 10 due to limited features. Data augmentation with DCGAN proves effective in improving accuracy in diverse regional contexts. Additionally, the study introduced Auxiliary Classifier GAN (ACGAN) as a comparative method, but its accuracy needs to improve due to uneven sample distribution and variations in generated sample quality. CNN-Custom's tailored design, filtering capability, and ensemble generation lead to stable and effective data augmentation results.

Conclusion

To summarize, the machine learning-based urban economic competitiveness model presented in this study offers a precise and reliable method for discerning regional disparities, aiding in the formulation of tailored policies and investment strategies. Potential applications include analyzing urban livability and sustainability with adjustments to feature indicators and experimental samples.

Furthermore, this research has broader implications for exploring regional differences at various scales, such as counties and villages. Future endeavors will involve integrating economic theories with neural networks and enhancing data feature engineering to advance artificial intelligence in urban economic research.