New Python Tool Maps Cause-and-Effect Chains in Complex Systems

Researchers have introduced DEMATELab, a Python-based desktop application that brings the full workflow of DEMATEL cause-and-effect analysis into a single graphical environment. The software, described in SoftwareX by Ali Torkayesh, Anahita Amiri and Sandra Venghaus, packages classical DEMATEL together with three uncertainty-aware extensions—triangular fuzzy, grey-number and Z-number-based variants—alongside threshold selection, clustering, visualization, export functions and an optional large language model module that drafts plain-language interpretations of the results.

DEMATEL, short for Decision-Making Trial and Evaluation Laboratory, uses matrix algebra and graph theory to identify and evaluate interdependencies among the components of a complex system, referred to as factors. In its standard form, experts assess the degree of influence between each pair of factors using qualitative linguistic terms such as “No Influence”, “Low Influence”, “Medium Influence”, “High Influence” and “Extreme Influence”, which are converted into numerical values of 0 through 4. These judgments form an initial influence matrix, which is aggregated across experts, normalized and transformed into a total influence matrix from which prominence and relation measures are computed. Factors with a positive relation value fall into the cause group, while those with negative values belong to the effect group.

The motivation for the new tool lies in a gap in the existing software landscape. According to the authors’ comparison of available implementations, PyDEMATEL supports classical and triangular fuzzy DEMATEL but does not cover interval grey numbers or Z-numbers, and it treats all experts as equally important. OnlineOutput is proprietary, subscription-based for large projects and limited to classical and fuzzy variants. The DEMATEL R package handles only the classical method and lacks a graphical interface, while pyDecision, RMCDA and JMcDM offer DEMATEL as one method within broader multi-criteria decision analysis libraries, without documented support for advanced extensions, clustering, threshold comparison or expert-data management.

DEMATELab addresses these shortcomings by integrating several solver variants within a unified framework, allowing users to address different types of uncertainty without switching tools. The software supports weighted aggregation of expert judgments, in which weights assigned to experts based on criteria such as experience, expertise or reliability must sum to one, rather than forcing every expert to contribute equally. Users can select a preferred thresholding technique, gain full access to all intermediate computations and matrices, and visualize cause-and-effect relationships through built-in causal diagrams and heatmaps.

The uncertainty extensions follow well-defined mathematical procedures. In fuzzy DEMATEL, each linguistic assessment becomes a triangular fuzzy number, and expert assessments are aggregated component-wise as weighted averages of the lower, middle and upper values. The total influence computation is performed separately for the three component matrices, and defuzzification uses the average of the three fuzzy values. Grey DEMATEL proceeds analogously with interval bounds, computing lower- and upper-bound matrices separately and applying interval arithmetic to prominence and relation. Z-number DEMATEL goes further by pairing each influence evaluation with a reliability rating—very low, low, medium, high or very high—so that both the ambiguity and the credibility of expert opinions are captured, with combined influence-certainty values drawn from the literature.

Threshold selection receives particular attention because the traditional approach, using the average of all entries in the total influence matrix, may not adequately capture the distribution of influence values. DEMATELab therefore offers alternatives including the first quartile, median and third quartile, which provide different levels of filtering intensity, as well as the Maximum Mean De-Entropy method, which seeks an optimal threshold by maximizing the informational content of the resulting network. Because threshold computation is separated from the solver layer, users can switch thresholding methods and regenerate causal diagrams without reconstructing the initial influence matrix—a task that previously required redoing all computations.

Beyond binary filtering, the software applies K-means clustering to the total influence matrix, grouping pairwise influence values into four clusters representing low, low-to-medium, medium-to-high and high influence. The authors note that thresholding alone discards many interactions and distinguishes only between retained and removed relationships, offering no insight into the intensity of influence between factor pairs. The clustering approach preserves this intensity information, allowing a more nuanced categorization of causal relationships.

The architecture is modular, with distinct layers for user interaction, analytical computation and utility services. The graphical interface generates dynamic input forms based on the chosen number of factors and experts, enforces fixed diagonal values and restricts data entry to drop-down lists to reduce invalid entries before computation begins. Judgments can be entered directly or imported from XLSX files, and results can be exported to spreadsheets or image files. An optional interpretation module constructs structured prompts for external LLM providers—OpenAI and Azure OpenAI with gpt-4o-mini by default, Gemini, OpenRouter and Groq—using the user’s own API key, which is not stored. The embedded prompt instructs the model to determine causal direction from relation values, treat prominence as importance rather than direction, and stay grounded in the computed data. The authors stress that a human-in-the-loop approach should be maintained to review and validate generated interpretations.

Validation was carried out with two published case studies. Using linguistic input from a fuzzy DEMATEL study on factors causing anchor loss in ships, which involved 18 factors and a single decision-maker, DEMATELab reproduced the same cause and effect groupings, and all exported tables matched the reported results. A second test applied grey DEMATEL to a study of 14 drivers of sustainable consumption and production adoption evaluated by two experts; the software again classified the same four drivers as effect drivers, and its causal diagram corresponded to the published one, with quantitative results differing only proportionally because of a different grey scale. The authors explicitly caution that validation was limited to the fuzzy and grey implementations, so these results should not be generalized to all four solver variants.

The authors also disclose that the software script was reviewed and optimized with the assistance of Anthropic’s Claude Sonnet model, with all outputs subsequently checked and verified, and that DeepL and Claude were used to improve the language of the manuscript itself. Among the acknowledged limitations, DEMATELab currently considers only four DEMATEL variants, uses K-means with a fixed four-cluster setting, and lacks project saving and loading; future versions are expected to add adjustable cluster counts, project-based workflows, improved LLM interpretation design and web-based deployment. Distributed through a public GitHub repository and installable via pip or from source, the tool aims to lower the barrier for researchers who lack the programming expertise to implement such analyses from scratch, while enabling comparative methodological studies and sensitivity analyses across different uncertainty representations.

Subject of Research: A Python-based software application for DEMATEL cause-and-effect analysis with multiple uncertainty models

Article Title: DEMATELab: A Python-based application for cause-and-effect analysis

Article References: Torkayesh, A., Amiri, A., & Venghaus, S. (2026). DEMATELab: A Python-based application for cause-and-effect analysis. SoftwareX, 36, Article 103172. https://doi.org/10.1016/j.softx.2026.103172

Image Credits: AI Generated

DOI: 10.1016/j.softx.2026.103172

Keywords: DEMATEL, cause-and-effect analysis, Python software, fuzzy DEMATEL, grey numbers, Z-numbers, multi-criteria decision analysis, K-means clustering, threshold selection, large language models, expert judgment, decision support