When evaluating complex, multi-dimensional systems—such as regional green growth, corporate credit risk, or digital economy maturity—researchers face a fundamental challenge: how should disparate indicators be aggregated into a single, comprehensive index?
If we want to measure the "green quality" of an industry, we might look at carbon emissions, waste management efficiency, and renewable energy investment. Assigning weights to these indicators traditionally relies on expert opinions or subjective frameworks like the Analytic Hierarchy Process (AHP). However, human judgment introduces bias and can fail when dealing with large, volatile datasets.
To eliminate this subjectivity, a prominent research direction leverages information entropy to calculate weights objectively based entirely on the underlying data's structure. Known widely as the Entropy Weight Method (EWM), this approach shifts the focus from what we think is important to how much information the data actually provides.
The Intuition: Information as Differentiation
The core mechanism of entropy-based weighting rests on a simple principle: an indicator is only useful if it helps us distinguish between the subjects we are evaluating.
Imagine evaluating the performance of several cities based on two metrics: air quality index (AQI) and the presence of municipal trash cans. If every single city has exactly 500 trash cans, this metric provides zero utility for ranking or comparing the cities. It contains no discriminatory power. Conversely, if the AQI varies wildly from city to city, that metric provides a wealth of information to help us differentiate them.
In information theory, Shannon entropy quantifies this uncertainty or disorder.
- High Entropy = High Uniformity: When data points for a given indicator are nearly identical across all samples, the data is highly ordered (low dispersion), meaning its entropy is high. Because it fails to differentiate the samples, it contains very little useful information for evaluation.
- Low Entropy = High Dispersion: When data points vary significantly across samples, the system is disordered (high dispersion), meaning its entropy is low. This variance provides a clear signal, making the indicator highly informative.
By calculating the entropy of each indicator, the system automatically assigns lower weights to uniform, uninformative metrics and higher weights to highly variable, informative metrics.
[Raw Data Matrix] ──> [Normalize Units] ──> [Calculate Entropy (E)] ──> [Compute Utility (1 - E)] ──> [Final Weights]
The Mathematical Mechanism
To understand why this works mathematically, consider the standard pipeline used across the literature, such as in work by Minjie Li et al. in their study of logistics green development (The Impact of Two-way FDI on GreenDevelopment Quality of the Logistics Industry– Based on the Threshold Effectof Environmental Regulation).
1. Normalization
Because indicators often use completely different units (e.g., metric tons of CO2 vs. GDP in USD), the data must first be normalized into a dimensionless scale between 0 and 1. For a positive indicator (where higher is better), the transformation is typically:
$$p_{ij} = \frac{x_{ij} - \min(x_j)}{\max(x_j) - \min(x_j)}$$
where $x_{ij}$ is the value of the $j$-th indicator for the $i$-th sample.
2. Calculating Information Entropy
Next, the normalized values are converted into a probability distribution for each indicator, and the information entropy $e_j$ is calculated:
$$e_j = -k \sum_{i=1}^n p_{ij} \ln(p_{ij})$$
where $k = \frac{1}{\ln(n)}$ is a constant that bounds $e_j$ between 0 and 1, ensuring comparability regardless of the sample size $n$.
3. Determining Information Utility and Weights
The "information utility" ($d_j$) of an indicator is simply the inversion of its entropy:
$$d_j = 1 - e_j$$
If all samples have identical values for an indicator, $e_j = 1$, meaning its utility $d_j = 0$. Finally, the objective weights ($w_j$) are derived by normalizing these utility values so that they sum to 1:
$$w_j = \frac{d_j}{\sum_{m=1}^m d_m}$$
This elegant formulation ensures that the data itself dictates the importance of each variable, completely bypassing human intervention.
Applications Across Domains
Because the entropy weight method is purely mathematical, it is agnostic to the field of study. Researchers have deployed it across a vast array of multi-criteria decision-making problems.
Environmental and Economic Synergy
A frequent application is measuring the balance between economic growth and environmental preservation. For instance, Qinfeng Xing applied the method to evaluate the coordination of low-carbon development and socio-economic growth in Anhui Province (Assessment of the coordination of low-carbon development and socio-economic development based on a comprehensive model: A case study of Anhui Province (China)), ensuring that the most distinct information drove the sustainability index. Similarly, Jianhong Cao et al. used information entropy weighting to synthesize multi-dimensional metrics like resource efficiency and pollution levels into a unified green growth index (Effect of Digital Financial Inclusion on Dredging the Path of Green Growth—New Evidence From Front-End and Back-End Perspectives).
Digital Economy and Infrastructure
The method is equally potent in quantifying abstract economic shifts. Qingran Guo et al. used the entropy method to aggregate diverse metrics like digital infrastructure and employment into a single representative score for the digital economy (The impact of digital economy and industrial structure distortion on Xinjiang’s energy intensity under the goal of “double carbon”).
Hybrid Systems: Correcting for Human Bias
While the pure entropy weight method provides complete objectivity, it has a notable limitation: it ignores the conceptual importance of an indicator. A highly volatile but trivial metric might receive a massive weight purely due to its statistical noise.
To solve this, modern research frequently pairs entropy weighting with other methodologies to create balanced, hybrid evaluation frameworks.
The AHP-Entropy Correction Layer
To ground statistical variance in real-world reality, researchers often combine subjective weights from the Analytic Hierarchy Process (AHP) with objective weights from entropy. Tong Wang demonstrated this in electronic banking risk evaluation (The Research of Electronic Banking Risk Evaluation Based on Comprehensive Assessment AHP-Entropy). By using entropy as a corrective layer, the model tempers human bias: if an expert claims an indicator is vital, but the data shows zero variance across samples, the entropy calculation automatically scales down its influence.
EWM-TOPSIS Integration
Another powerful combination is pairing EWM with the Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS). In this setup, EWM determines the objective weights of the criteria, and TOPSIS calculates the geometric distance of each alternative from an ideal "best" and "worst" solution.
Zhenlin Dai et al. successfully utilized this hybrid EWM-TOPSIS model to map light pollution levels across diverse urban and rural landscapes in Beijing (Research on light pollution level based on comprehensive evaluation —— A case study of Beijing). By relying on data dispersion to weight satellite and socio-economic imagery, the authors ensured a robust comparison across geographically and economically diverse regions without researcher bias.
Multi-Objective Optimization
Beyond simple ranking, entropy weighting is also used to balance competing goals in engineering. Jing Qiu and Renjie Mao integrated information entropy into a particle swarm optimization algorithm for renewable energy power grids (Optimized Particle Swarm Algorithm for Advanced Bi-Level Dispatch in New Energy Power Systems). By mathematically weighting competing objective functions—such as minimizing operational costs versus maximizing grid stability—the authors let the information density of the grid variables drive the optimization, resulting in a highly stable, data-driven dispatch strategy.