AI and fuzzy framework reconstructs China’s economic growth quality index

Researchers have built a hybrid artificial intelligence and fuzzy algorithm framework that reconstructs a multidimensional Economic Growth Quality Index from provincial-level data covering China from 1990 to 2021, reporting low error values and strong consistency in reproducing the composite index. The study, published in Discover Computing, combines objective statistical weighting, expert-informed fuzzy weighting, and a deep neural network whose settings are tuned by a nature-inspired optimization algorithm.

The work responds to a recognized shortcoming in how economic growth quality is measured. Conventional approaches, the authors argue, do not effectively bring together multidimensional indicators of regional growth, uncertainty in subjective weightings, and non-linear computational relationships within a single framework. Their proposed system, called AI-FAF, is designed to integrate all three elements in one end-to-end computational pipeline rather than treating weighting, uncertainty handling, and prediction as separate tasks.

The framework rests on three cooperating components. First, the Entropy Weight Method derives objective weights for each indicator from the information dispersion in the data, so that indicators with greater variability across provinces and years receive higher statistical importance. Second, a Fuzzy Analytic Hierarchy Process incorporates expert judgments under uncertainty, using fuzzy pairwise comparisons and linguistic variables so that strategically or politically relevant indicators are not underestimated simply because their statistical dispersion is low. The two weighting schemes are fused into a single hybrid weight vector, with a parameter controlling the balance between data-driven and expert-driven importance.

Third, the weighted indicators feed into an Enriched Deep Neural Network, which learns non-linear relationships among the economic, social, technological, environmental, public-investment, and innovation variables and produces a continuous estimate of the index. A Bald Eagle Search metaheuristic, inspired by the hunting behavior of bald eagles through selection, spiraling, and swooping phases, is used exclusively to optimize the network’s hyperparameters such as learning rate and hidden layer configuration, rather than to make predictions itself.

The underlying dataset comes from Kaggle and covers regional development patterns at the provincial level in China between 1990 and 2021. Thirteen indicators were selected based on their availability and relevance: gross domestic product, disposable income, healthcare growth, digital penetration, urbanization rate, labor structure, fine particulate matter concentration, carbon emissions, government spending, research output density, intellectual property rights, green productivity, and research and development expenditure. Together these reflect household welfare, structural and social development, technological progress, environmental sustainability, public sector investment, and innovation capacity.

Data preparation was deliberately structured to prevent information leakage. The panel was split chronologically into 80 percent training and 20 percent testing sets, and all preprocessing steps, including min-max normalization and linear interpolation for missing values, were fitted on the training data alone before being applied unchanged to the test set. Five-fold cross-validation was performed exclusively within the training subset, with temporally ordered folds, while the held-out test set was used only once for final independent evaluation. The authors report that the framework, implemented in Python 3.11, achieved a Mean Absolute Error of 0.023 and a Root Mean Square Error of 0.0346, alongside a Mean Squared Error of 0.0012, a Mean Absolute Percentage Error of 2.87 percent, and a coefficient of determination of 0.96.

An important caveat runs through the paper: because the Economic Growth Quality Index is constructed from the same thirteen indicators that are fed into the model, the task is fundamentally one of non-linear reconstruction of a predefined composite index, not independent prediction of an externally measurable economic outcome. The reported accuracy of 97.81 percent, which is an error-based percentage measure rather than a classification rate, should therefore be interpreted as a measure of estimation consistency and model agreement with the EWM-FAHP-derived index, not as evidence of causal relationships or genuine forecasting capability. The authors explicitly note that independent forecasting using lagged indicators and temporally separated outcomes remains a direction for future validation.

Ablation experiments illustrate how each module contributes. An EDNN-only baseline achieved an EGQI prediction measure of 91.84 percent. The framework was also benchmarked against retrained baseline models, including Random Forest, XGBoost, Support Vector Regression, Huber Regressor, Decision Tree, Multilayer Perceptron, LSTM, and ARIMA, all retrained on the same dataset with identical indicators, preprocessing, and data splits, with AI-FAF showing better performance on the reported error metrics within that constrained experiment.

Substantively, the analysis points to pronounced regional differences in growth quality, with innovation capability and environmental sustainability emerging as key factors. Regression diagnostics using standardized coefficients indicated positive associations for GDP (0.421), disposable income (0.318), research and development expenditure (0.294), and green productivity (0.252), alongside negative associations for PM2.5 concentration (0.196) and carbon emissions (0.142), though the authors caution that because the index is built from the same indicators, these results provide sensitivity information rather than independent causal validation. Exploratory visualizations traced rising digital penetration and research spending after 2010, carbon emissions peaking near 0.95 on a normalized scale before easing to 0.85, and top-performing provinces improving from index values of roughly 0.31 to 0.47 up to 0.76 to 0.90, while the bottom five provinces rose only from 0.10 to 0.23 to 0.41 to 0.58, underscoring persistent inequality across regions.

The authors position the framework as a decision-support tool that could help policymakers identify which dimensions of regional development need attention, for example by flagging weak environmental performance or low innovation capacity in a given province’s index profile. However, they are careful to state that policy implications drawn from the model have not been empirically validated through real-world policy actions, and that the framework should be construed as an analytical assessment tool rather than one demonstrated to produce policy effectiveness. Limitations include dependence on the quality and completeness of the available panel data, confinement to a single country and indicator set, and the inability to test generalization across other regions, datasets, or economic systems. Future work, they suggest, should explore lagged variables, longer and higher-frequency datasets, explainable AI methods, and validation across countries to establish genuine forecasting viability and broader applicability.

Subject of Research: A hybrid AI and fuzzy algorithm framework for assessing economic growth quality using Chinese provincial panel data

Article Title: AI and fuzzy algorithm-driven comprehensive assessment for economic growth quality

Article References: Tsao, S., & Chen, W. (2026). AI and fuzzy algorithm-driven comprehensive assessment for economic growth quality. Discover Computing, 29(1), Article 817. https://doi.org/10.1007/s10791-026-10724-7

Image Credits: AI Generated

DOI: 10.1007/s10791-026-10724-7

Keywords: economic growth quality, artificial intelligence, fuzzy algorithms, deep neural network, entropy weight method, China provincial data, composite index, machine learning, Bald Eagle Search, economic indicators, sustainability, regional development