TAAA Digital → Text Analyzer

TAAA from Digital → Text with mapping power(MP=kappa) vs True_Label

Run(Lable item testlet names, Assign themes to person, and then compute MP between themes and latent traits

Dummy patient values (comma-separated; same order as item columns). Leave empty → use top-1 row.

Detected item column order (from uploaded CSV/TSV/TXT header):

Status: Ready.

After you run, the report will appear here.

Install dependencies:
pip install -r requirements.txt
Start server:
python main.py
Open browser:
http://127.0.0.1:8000/
Click Run (Upload) or Run Demo. The report will appear in the preview iframe.

If pred_seq is empty → dummy patient = copy of the top-1 row (training data unchanged).
If pred_seq is provided → dummy patient uses your values (comma-separated, same order as item columns).
If some values are blank (e.g., 80,,60,...) → blanks will be filled by feature means (so the run can continue).

Each run generates a new folder (example):
web_reports/<run_id>/.../report.html
If something fails, click the “Log” link shown after running to see traceback/subprocess output.

Tip: If backend returns HTML error pages (HTTP 500), this page will show that HTML directly in Debug instead of crashing on JSON parsing.

The analysis generates 6 figures + 2 tables, all bundled into a single report.html.

Kano (Person + Item, interactive HTML)
Volcano (interactive HTML + PNG; up to 500 features; PSA is always shown even if non-significant)
Heatmap (rows=item, cols=person; columns grouped by TRUE_LABEL; item labels use p<0.001 / p=0.xxx)
Sankey (interactive HTML + SankeyMATIC code + clickable SankeyMATIC link)
One-link network (maximum spanning tree on core items)
ROC combined (three curves: TRUE vs SS(i), theme2 vs SS(i), TRUE vs theme2)
Table 1: 2×2 confusion matrix + Accuracy / Precision / Recall / F1 / Specificity
Table 2: Varimax-rotated 2-factor loadings (two heatmap item clusters) + item SS(i) + correlation statistics

ℹ️ Person relations（WCD / one-link edges）計算規則說明（點擊展開）

Euclidean distance → 關聯強度（WCD）
對任兩個 person（vertices）計算 Euclidean distance d，並轉為 0–1 的關聯強度：
WCD = (max(d) - d) / (max(d) - min(d))
數值愈大表示兩者關聯愈強。
pair 標準化、排序與去重
每一對關係表示為 (term1, term2)，其中：
alphabetic 較小者為 term1，較大者為 term2，以避免重複 pair。

若仍可能出現同分，加入微擾的 tie-break：
WCD' = WCD + nrow(relations) - row_index
確保排序時不會完全同分。
one-link 規則
以 term2 為分組鍵，只保留該 term2 對應 WCD（或 WCD'）最大 的一筆關係。
保留下來的邊即構成 one-link edges。
Top 20 vertices 限制
最終 one-link edges 的兩端點（term1, term2）必須皆屬於 Top 20 vertices；否則不納入最終 relations。