Bibliometric Analysis

Bibliometric Analysis

Turn a research topic into a defensible map of its scholarly literature: a screened corpus dataset, a structured metrics file, a full chart suite, and a written report with research-gap implications.

The reference implementation (in scripts/ of this directory) is Python, but the workflow is language-agnostic: every stage is defined by its inputs, outputs, and rules — implement each stage in whatever stack the user prefers (Python, R, JavaScript/TypeScript, Julia, shell + jq, …). Treat scripts/*.py and run_pipeline.sh as an executable specification, not a requirement.

Ground rules

1. Stages communicate only through files — a coding-sheet CSV and an insights JSON. Keep the schemas below byte-compatible and any stage can be rewritten or replaced in another language without touching the others.
2. One self-contained output directory per run: reports/<YYYYMMDD_HHMM>_<topic-slug>/ holding the coding sheet, insights JSON, and all charts. Never scatter run artifacts across shared folders.
3. Fail fast: run stages in order; if a stage errors, stop and report rather than continuing with partial data.
4. Coding-sheet language: column names and category labels below are the canonical contract (in Indonesian, inherited from the reference corpus — e.g. lembar koding = coding sheet). Chart titles/axes may be in the user's report language; do not rename the CSV columns without migrating every downstream stage.

Stage 0 — Scope the study

Before any code, pin down and write into the run directory as study_design.md:

• Topic and angle (e.g. "digital activism", "ESG perception", "hoax detection").
• Time window and document filter (reference default: articles from 2010 onward, abstract required).
• Geographic/contextual focus, if any (reference: Indonesia).
• Query set: 5–10 search queries covering synonyms, adjacent framings, and — if the field publishes in one — the local language. Reference example:
  • "digital activism" AND ("Indonesia" OR "Indonesian")
  • "hashtag activism" AND ("Indonesia" OR "Indonesian")
  • "aktivisme digital" AND "Indonesia" (local-language variant)
• Domain adaptation rules to swap in (see checklist at the end): relevance signals, platform list, theme/theory keyword rules, stop-keywords.

Stage 1 — Collect (OpenAlex API)

Output: lembar_koding_bibliometrika.csv — the canonical coding sheet.

API contract

• Endpoint: GET https://api.openalex.org/works
• Parameters:
  • search=<query> (one request series per query in the query set)
  • filter=from_publication_date:<YYYY>-01-01,type:article,has_abstract:true
  • per-page=100, page=1..N
  • select=id,doi,title,authorships,publication_year,primary_location,biblio,cited_by_count,keywords,concepts,language,abstract_inverted_index,referenced_works
• Politeness: send a User-Agent identifying the project with a mailto: address (OpenAlex "polite pool") — the reference fetcher takes --email (or the OPENALEX_EMAIL env var via run_pipeline.sh) — and sleep ~0.25 s between paged requests.
• Retry: on HTTP 429 retry up to 4 times with exponential backoff (wait 2^attempt + 2 seconds). On other errors, abandon that page and move on.
• Stop paginating a query when a page returns fewer results than per-page.
• Caps (defaults; size to the field): 10 pages per query, 500 final rows (MAX_ROWS).

Batch metadata resolution (used again in Stage 4)

To label referenced works (co-citation network, reference genealogy): GET https://api.openalex.org/works?filter=openalex:W1|W2|W3&select=id,title,authorships,publication_year (pipe-separated OR of OpenAlex work IDs, ≤50 per call). Label format: LastName (Year).

Record processing rules

• Deduplicate across queries by OpenAlex work ID.
• Sort by citation count descending, then publication year descending; keep the top MAX_ROWS.
• Reconstruct abstracts from abstract_inverted_index: it maps each word to its positions — flatten to (position, word) pairs, sort by position, join with spaces.
• Authors: Last, First format, first 4 authors, then et al.; ;-separated.
• Collaboration type from authorships' institution country codes: 1 author → Tunggal; >1 country → Multi-negara; 1 country → Satu-negara; no country data → Multi-penulis (afiliasi tidak diketahui).
• Citations per year = cited_by_count / max(current_year − publication_year, 1).
• Platform studied — keyword match on title (first match group per platform): twitter|tweet|x.com → Twitter/X; facebook; instagram; youtube; tiktok; whatsapp|telegram|line → WhatsApp/Telegram/Line; reddit|discord → Reddit/Discord. No match → Tidak spesifik / multi-platform.
• Theme / theory inference from the OpenAlex concept hierarchy. Note: OpenAlex is deprecating concepts in favor of topics (field → subfield → topic) — if concepts is absent, anchor inference on topics levels instead. Any equivalent leveled taxonomy works:
  • Tema_Utama = first level-2 concept; fallback first level-1.
  • Subfield = first level-1 concept.
  • Teori_Kerangka = first level-3 concept, else "unspecified".
  • Metode_Penelitian = up to 2 concepts whose names match method-ish keywords (analysis, method, model, algorithm, technique, approach, framework, survey, experiment, study, evaluation, review, mapping, mining, network, simulation).

Canonical coding-sheet schema

Write UTF-8 with BOM so Excel opens it cleanly (Python calls this encoding utf-8-sig; in R, readr::write_excel_csv()).

Stage 2 — Screen (automated inclusion/exclusion)

Input/output: the coding sheet — update Kriteria_Inklusi / Alasan_Eksklusi in place (default every row to Dimasukkan).

Apply, in order:

1. Duplicates: normalize titles to lowercase alphanumerics only; second and later occurrences → exclude, reason Duplikasi.
2. Abstract adequacy: fewer than 30 words → exclude (incomplete records, editorials, non-research items).
3. Topical relevance (only when the study has a geographic/contextual focus): require signals of the focus context in title/abstract (reference used tokens like jakarta, jokowi, kpk). Also drop papers where the context is only a marginal data point — operational test: no context signal in the title AND the context country absent from author affiliations AND the abstract name-drops it alongside several other countries (cross-national survey with a one-line Indonesia result).

Print a screening summary: totals included/excluded and a breakdown by exclusion reason. Keep excluded rows in the file (audit trail), never delete them.

Stage 3 — Extract insights (JSON)

Input: only rows with Kriteria_Inklusi == Dimasukkan. Output: insights_data.json — one key per indicator, each a ranked {label: count} map or per-year structure. This file is the numeric backbone the report quotes verbatim; every number in the report must be traceable to it.

Indicators (reference keys):

• tren_publikasi — articles per year.
• distribusi_tema / distribusi_tema_detail — theme counts (raw / after breaking generic buckets apart).
• asal_peneliti, kolaborasi, top_institusi, top_jurnal, top_negara.
• top_keywords — from author + index keywords: split on ;/,, lowercase, collapse whitespace, drop tokens <3 or ≥50 chars, drop stop-keywords (domain-generic terms like indonesia, social media, digital, study, analysis, based, network — anything that would top every chart without informing), dedupe per article, count.
• kolaborasi_negara — for each multi-country article, count all sorted country pairs.
• sitasi — per year: {total_articles, total_citations, avg_citations}.
• distribusi_teori, distribusi_teori_detail, evolusi_teori (theory counts per year, keep theories with ≥5 articles).
• teori_vs_metode and tema_vs_metode — sparse {"Theme | Method": count} pairs.
• dampak_sitasi_per_tema / _kolaborasi / _bahasa / _penerbit — avg, n, total citations per group (require n ≥ 2).
• distribusi_penerbit — publisher tier by DOI prefix. Reference prefix map: 10.1080 Taylor & Francis, 10.3390 MDPI, 10.1177 SAGE, 10.1007 Springer, 10.1017 Cambridge UP, 10.1093 Oxford UP, 10.1016 Elsevier, 10.1002/10.1111 Wiley, 10.1186 BMC, 10.3389 Frontiers, 10.1145 ACM, 10.1109 IEEE, 10.2991 Atlantis Press (→ proceedings). Other 10.x → other journal; no DOI → Tanpa DOI. Extend with the local publishers your field actually uses.

Stage 4 — Visualize

Input: screened coding sheet (+ live OpenAlex metadata calls for co-citation labels). Output: PNG charts (reference: 150 dpi, ~11–16 in wide) in the run directory. Every chart states its n and thresholds in the title or caption.

Chart catalog

Visual style (academic, print-safe)

• Serif font family (STIX / Times / DejaVu Serif), base size 11, bold titles.
• No top/right spines; dashed light-gray grid behind data; white background.
• Muted palette, e.g. navy #2B4C7E, sage #5B8C5A, brick #A45A52, steel #6B7AA1, tan #C49A6C, warm gray #7A6C5D (warm tones for inferred categories, cool for original — legend explains the distinction).
• Always label values on bars; wrap long category labels onto 2–3 lines.
• Networks: force-directed layout with a fixed random seed (reproducibility), remove isolated nodes, white node outlines.

Network construction recipe (any language)

Count pair co-occurrences with the sorted-pairs trick: for each article take its deduplicated item set, generate all 2-combinations of the sorted set, increment pair counters. Filter by frequency/co-occurrence thresholds, cap node count, lay out with a force-directed algorithm (e.g. Fruchterman–Reingold / spring layout).

Stage 5 — Report

Inputs: insights JSON + charts + (if provided) the user's research proposal. Outputs: an analytic-memos document (visualizations_insights.md) and/or a findings chapter. Write the report in the user's working language even though the coding-sheet labels are Indonesian — quote labels as-is when citing the data.

Rules learned the hard way — follow them:

1. Never let a generic category dominate. If one bucket like "Digital activism (umum)" or "unspecified theory" tops a distribution, break it down with keyword rules over title/keywords/abstract before charting and reporting (the reference corpus resolved 93% of its generic bucket this way). Same for "local researchers" — resolve to institutions.
2. One analytic memo per chart: Key insight → interpretation → historical / contextual explanation (e.g. a 2019 publication spike tied to real-world protest waves). Numbers must come from the insights JSON, not eyeballed from charts.
3. Separate diagnostic charts from report charts. Citation-impact-by-collaboration, publisher distribution, platform evolution, and the theme×/theory×method matrices are exploratory; exclude them from formal reports unless they answer a stated research question.
4. Align with the research paradigm stated in the proposal (epistemology, methodological framing) — findings chapters must speak its language, at the qualification level the user is working at.
5. End with research gaps: sparse themes, missing platforms, absent regions or collaboration patterns are the deliverable — they justify the next study.

Orchestration

Run order: collect → screen → insights → visualize → report, each stage reading the previous stage's files from the shared run directory. Reference runner: run_pipeline.sh [keywords…] — builds the timestamped run directory (under the directory you invoke it from), chains the stages, stops on first failure. Re-running a single stage against an existing run directory must work (all stages take the run directory as their target).

Choosing a stack

Any combination providing: an HTTP client, CSV + JSON I/O, a plotting library, and a force-directed graph layout. Suggested equivalents per stage:

Prefer the ecosystem the user's repo already uses; when unsure, default to the reference Python stack so scripts/ can be reused as-is (third-party deps are listed in scripts/requirements.txt; fetch/screen/insights are stdlib-only).

Domain adaptation checklist

When applying this to a field other than the reference (digital activism in Indonesia), replace:

• Query set (synonyms + local-language variant)
• Relevance signals for the contextual focus (Stage 2, rule 3)
• Platform list (Stage 1 inference) — e.g. finance apps, health platforms, LLM tools
• Theme/theory keyword rules for generic-bucket breakdown (Stage 5, rule 1)
• Stop-keyword list (terms that are generic within this field)
• Country-code focus and publisher prefix map, if the field publishes elsewhere
• Time window start (2010 fits social-media topics; older fields need more)