DEMATELab — Cause-and-Effect Analysis Toolkit

DEMATEL (Decision-Making Trial and Evaluation Laboratory) is a structural modelling method for untangling cause-and-effect relationships between factors. In practice it usually lives in fragile, hand-built spreadsheets that break down the moment you add a second expert or switch from crisp numbers to fuzzy or grey ones. DEMATELab turns the whole pipeline — linguistic input, expert weighting, threshold selection, the total-relation matrix, and clustering — into one desktop GUI, with four solver families behind a single interface.

Problem

A real DEMATEL study rarely uses one expert and one number scheme. Researchers need to:

• Collect judgments from several experts and weight them differently.
• Choose a number representation — crisp, fuzzy, grey, or Z-number — to model uncertainty in those judgments.
• Pick a threshold method to decide which relationships in the total-relation matrix are significant enough to keep.

Doing this by hand means re-deriving the matrix algebra for each variant and re-doing it every time an expert's input changes. The math is identical in shape but different enough per variant (fuzzy arithmetic, grey bounds, Z-number reliability) that spreadsheets become error-prone and unauditable.

Approach

The app is organised as a linear workflow — Setup → Input Forms → Matrix Editor → Run Analysis → Results — so the method's steps map directly onto the UI.

Four solver types share one interface, each handling its own arithmetic:

• Classical DEMATEL — crisp 0–4 influence scale.
• Fuzzy DEMATEL — triangular fuzzy numbers (l, m, u).
• Grey DEMATEL — grey number intervals [l, u].
• Fuzzy Z-DEMATEL — combined influence–reliability judgments (e.g. "High influence with M reliability"), where reliability runs VL, L, M, H, VH.

Five threshold methods decide which influences survive into the causal diagram: Average, First Quartile, Median, Third Quartile, and Maximum Mean De-Entropy (MMDE) — the last computing the cut-off from the information content of the matrix rather than a fixed percentile.

Multi-expert support lets each expert have a name and weight; the matrix editor presents linguistic dropdowns per expert so judgments are entered in words, not raw numbers.

Results are split across tabs — Summary, Matrices, R/C scores (prominence and relation), Causal Diagram, and Heatmap — plus K-Means clustering over the total-relation matrix to group factors. Everything exports to XLSX, and visualizations export to PNG.

AI explanation is optional: the AI tab sends the computed results to an LLM (Gemini, OpenAI, OpenRouter, or Groq) and returns an academic-style narrative interpretation. API keys are held in memory for the session only and never written to disk.

Linguistic Scales

The same linguistic label maps to a different numeric representation depending on the chosen solver:

Fuzzy Z-DEMATEL layers a reliability term on top of these influence judgments.

What It Delivers

• 4 solver families (Classical, Fuzzy, Grey, Fuzzy Z) behind one consistent workflow.
• 5 threshold methods, including MMDE.
• Multi-expert aggregation with per-expert weights.
• 5 result views plus K-Means cluster analysis.
• XLSX + PNG export and optional LLM interpretation across 4 engines.
• Cross-platform: Windows, macOS, Linux; installable via pip install git+….

Tradeoffs

• Desktop, not web. PySide6 gives a responsive native matrix editor and rich plotting without shipping a backend — at the cost of a pip install instead of a URL. For a research tool used locally with private expert data, that was the right call.
• In-memory API keys. Keys are never persisted, which is safer but means re-entering them each session. Acceptable for an analysis tool that isn't run continuously.

What I'd Do Differently

• Add a saved-project format so an in-progress study (experts, matrices, settings) survives closing the app, not just an export.
• Factor the per-solver arithmetic behind a single solver interface earlier — the variants share enough structure that a stricter abstraction would cut duplication.

Known Limitations

• No automated test suite ships with the project yet; correctness rests on the matrix algebra and manual validation against known DEMATEL results.
• AI interpretation requires the user's own API key and network access.
• Grey and Z-number variants assume the fixed linguistic scales above rather than user-defined ones.

Architecture

Setup panel  (solver type + threshold method + #experts + #factors)
   └─ Input Forms  (factor names, expert names + weights)
        └─ Matrix Editor  (linguistic dropdowns, one matrix per expert)
             └─ Solver  (Classical | Fuzzy | Grey | Fuzzy-Z)
                  ├─ aggregate experts by weight  → direct relation matrix
                  ├─ normalize → total relation matrix (T)
                  ├─ R / C vectors  (row + column sums of T)
                  ├─ threshold (Average | Q1 | Median | Q3 | MMDE)
                  └─ K-Means clustering over T
                       └─ Results tabs  (Summary, Matrices, R/C, Causal, Heatmap)
                            ├─ export → XLSX / PNG
                            └─ AI tab → LLM → narrative interpretation

Stack