TAAA without mapping power(MP, i.e., no True_Label in the last column)

ReadMe — Methods (A–E)

Notes: Bootstrapping & OOL (C/D) are applied only to Type-1 text co-word inputs (rows=articles). For numeric Type-2/3/4 matrices, C/D is skipped even if enabled.

A — Inter/ Intra-rater (core-theme sample)

• Randomly sample a fraction of articles from the major themes (core themes).
• Compare TAAA vs Reviewer 1, TAAA vs Reviewer 2, and Reviewer 1 vs Reviewer 2.
• Outputs: confusion matrices + Cohen’s κ (kappa).

B — External generalization (optional)

• If your data contains TRUE_LABEL (external ground truth), TAAA themes are compared to it.
• Outputs: confusion matrix + Cohen’s κ.

C — Bootstrapping stability (optional)

• Resample the sampled articles with replacement for bootstrap_reps times.
• Recompute κ to quantify stability; provide mean and percentile CI (2.5/97.5).

D — OOL (holdout) robustness (optional)

• Repeatedly draw random holdout subsets (size = holdout_frac) for ool_reps times.
• Compute κ on the holdout subset to quantify robustness against dataset shift.

E — Sankey (Theme → Term flow)

• Create SankeyMATIC-compatible text describing flows from theme leaders to terms (top nodes only).
• Outputs: terms_sankeymatic.txt and a ready-to-open SankeyMATIC URL in the report.

How it works (2 stages)

Stage 1 — Auto-classify input data (4 types)

1. Type 1: text in (almost) all cells (keyword/phrase occurrences per row)
2. Type 2: numeric square matrix (k×k) with colnames == rownames
3. Type 3: numeric square matrix (k×k) with colnames != rownames
4. Type 4: numeric rectangular matrix (m×n) with colnames != rownames

Stage 2 — Convert to edges + nodes for FLCA

• Type 1 → build co-occurrence edges: (term1, term2, WCD)
• Type 2 → edges from upper triangular: WCD = matrix[i,j]
• Type 3 & 4 → edges from all numeric cells: term1=rowname, term2=colname, WCD=value

Node table for FLCA

• name
• value = weighted degree (sum WCD)
• value2 = degree count (# non-zero connections) — matrix mode uses value2=value