Researchers have unveiled a new artificial intelligence framework that combines a large-scale knowledge graph with contrastive learning to evaluate how developed a digital economy is, reporting markedly higher accuracy than conventional statistical and machine learning approaches. The algorithm, named KGC-DEE, was described in a study published in Discover Artificial Intelligence, and it achieved an accuracy of 0.923 and an F1-Score of 0.918 on a self-built benchmark dataset, while reducing root mean square error and mean absolute error to 0.068 and 0.052 respectively. The work addresses a long-standing problem in economic measurement: how to capture the tangled, nonlinear web of relationships among regions, companies, technologies and policies that determines whether a digital economy is thriving.
Traditional approaches to assessing digital economic development have relied heavily on macro-statistical indicators such as internet penetration rates, the share of the information and communications industry in gross domestic product, and e-commerce transaction volumes. Analysts typically build a multidimensional index system, assign weights using techniques like the entropy weight method or the analytic hierarchy process, and compute a composite score for comparison across regions or over time. These methods are popular with policymakers because the underlying data are widely available and the calculations are transparent. However, the study’s author argues that they struggle to model the complex nonlinear relationships and hidden correlations among the many entities involved in a digital economy, which can leave the resulting evaluations at a superficial level.
Machine learning brought improvements. Support vector machines have been used to classify regional levels of digital development, while ensemble methods such as random forests and gradient-boosted decision trees can fit complicated mappings between economic indicators and development outcomes. Yet even these data-driven models generally treat each evaluation sample as an independent data point, ignoring the structured information embedded in the economic connections and technology diffusion effects that link regions together. Because a region’s digitalisation level is influenced by affiliated enterprises and policies in neighboring areas, that independence assumption rarely holds in a real economic system, limiting the depth and insight of the resulting assessments.
The new framework rests on two complementary technologies. A knowledge graph is a semantic network that integrates information in structured form, and in this case it links five core entity types: regions, industries, technologies, enterprises and policies. Relationships such as ownership, dependence, support and location describe how these entities interact. To build the graph, the researcher fused data from the National Bureau of Statistics, annual reports of listed companies, industry white papers, patent databases and government policy texts, using BERT-based entity identification and rule-based relation extraction. The resulting network covers approximately 120,000 entities and 650,000 relationships spanning 2010 to 2023, forming a rich system of prior knowledge about the digital economic ecosystem.
On its own, however, a knowledge graph’s embeddings can be sensitive to noise and insufficiently discriminative. That is where contrastive learning enters. This self-supervised technique learns high-quality features without manual labelling by pulling the representations of similar samples closer together and pushing dissimilar ones apart. The framework uses a graph attention network as its encoder, allowing the model to weigh the differing importance of neighbouring nodes, and generates two augmented views of the graph through two enhancement strategies: relational path discarding, which randomly removes edges to simulate missing links, and feature masking, which randomly zeroes out feature dimensions so the model cannot over-rely on a few attributes. Each node’s embeddings from the two views form a positive pair, optimised through a normalised temperature-scaled cross-entropy loss.
The final evaluation stage focuses on region nodes. An optimised embedding for each region is passed through a multi-layer perceptron with a hidden layer, dropout for regularisation, and a sigmoid output that constrains the predicted score between zero and one. Training minimises the mean squared error between predicted scores and expert-assigned labels. Crucially, the framework is trained end-to-end with a multi-task joint loss that combines knowledge graph embedding loss, contrastive learning loss and evaluation loss, each weighted by a tunable hyperparameter, using alternating optimisation and the Adam optimizer until convergence.
Experiments were conducted on the self-built DigitalEconomy-2024 dataset, which aggregates 10,000 samples of economic indicators, policy impacts and technological innovation dimensions from 2018 to 2023. Missing values were handled with mean imputation and matrix factorisation, and outliers were filtered using the interquartile range method. Labels were assigned by a panel of five independent domain experts through a multi-round Delphi method, categorising development levels as high, medium or low, with labelling consistency verified by Fleiss’ kappa at 0.82, indicating strong agreement. The data were split 70-15-15 into training, validation and test sets, and all models were implemented in PyTorch on an NVIDIA RTX 3090 GPU.
The results showed KGC-DEE outperforming every baseline tested, including linear regression, random forests, support vector machines, the graph neural network GCN and the graph contrastive learning method GraphCL. Compared with traditional machine learning methods, improvements ranged from 11.3 to 15.4 percent across the metrics. Against graph-based baselines, the framework delivered a relative accuracy gain of 11.1 percent over GCN and 6.5 percent over GraphCL, and F1-Score gains of 11.7 and 7.5 percent respectively. An ablation study underscored the synergy at the heart of the design: removing the knowledge graph cut accuracy from 0.923 to 0.832 and raised RMSE from 0.068 to 0.108, while removing contrastive learning reduced accuracy by 7.4 percent, and removing both components caused a sharp 15.4 percent performance drop.
The study also probed the model’s inner workings. A case analysis centred on the city of Hangzhou showed how the knowledge graph encodes explicit relationship chains linking enterprises, technologies, industries and policies, with policy support relationships accounting for 35 percent and technology domain associations reaching 60 percent of the visualised subgraph, providing interpretable structured support for the evaluation. Training curves demonstrated that the contrastive loss fell from an initial 2.8 to 0.5 within 50 epochs, with validation loss converging to 0.52. Hyperparameter analysis found peak performance when the contrastive learning weight was 0.5 and the evaluation loss weight was 0.4, while a temperature parameter of 1.0 yielded the most stable convergence; setting it to 0.1 slowed convergence by 25 percent and increased residual oscillation by 18 percent. A scatter plot of predicted versus expert scores showed a strong positive correlation, with 78 percent of predictions correct and a mean squared error below 0.05.
The author acknowledges important limitations. Expert labelling, even with the Delphi method, carries a degree of subjective bias, and because the dataset primarily reflects the Chinese economic context, the model’s generalisability to other national or regional systems with different industrial structures remains unverified. The framework also depends heavily on the quality and completeness of the underlying knowledge graph, since erroneous entity-relation extractions can propagate noise into the contrastive learning module and skew assessments. In addition, the multi-task training of graph attention networks and contrastive learning on large heterogeneous graphs is computationally expensive, which may limit real-time use in resource-constrained settings. Future work, the paper notes, will explore weakly supervised labelling to reduce human bias, cross-lingual and cross-regional knowledge graphs to improve generalisability, and lightweight graph neural network architectures to cut computational costs.
Subject of Research: A knowledge graph and contrastive learning framework for evaluating digital economy development
Article Title: A knowledge graph and contrastive learning framework for evaluating digital economy development
Article References: Li, C. (2026). A knowledge graph and contrastive learning framework for evaluating digital economy development. Discover Artificial Intelligence, 6(1), Article 982. https://doi.org/10.1007/s44163-026-02074-6
Image Credits: AI Generated
DOI: 10.1007/s44163-026-02074-6
Keywords: knowledge graph, contrastive learning, digital economy, graph attention network, machine learning, economic evaluation, self-supervised learning, TransH, GraphCL, benchmark dataset, Delphi method, representation learning
News Source: Denise Maddox. (October 28, 2026). Knowledge Graph Meets Contrastive Learning to Score Digital Economy Growth. Scienmag.
