thaieda 1.6.0

AutoEDA สำหรับข้อมูลภาษาไทย — Exploratory data analysis that speaks Thai

ThaiEDA

Exploratory data analysis that actually understands Thai.

---

What is ThaiEDA?

ThaiEDA is a Python library that automates exploratory data analysis for Thai and mixed Thai/English datasets. You give it a DataFrame, it gives you back a full report — smart pre-analysis, language detection, column types, quality issues, anomalies, cross-column insights, charts, and an executive-style HTML report. All in one line.

It handles the things generic EDA tools miss: Buddhist Era dates, Thai numerals, zero-width spaces, Thai vowel/tone marks, mixed Thai/English cells, mojibake encoding, Thai month names, national ID card validation, Thai address parsing, and PII like phone numbers.

---

Quick Start

pip install thaieda

import thaieda
import pandas as pd

df = pd.read_csv("data.csv")
result = thaieda.run(df)          # full EDA in one line
result.to_html("report.html")     # self-contained HTML report

pip install thaieda ติดตั้งทุกอย่างเลย — Thai tokenizer, NER, ML, Excel, stats, encoding detection, interactive charts ไม่ต้องใส่ extras

---

Why ThaiEDA?

Generic tools don't understand Thai data. Pandas Profiling, ydata-profiling, and Sweetviz are great — until you feed them Thai data. They miss Buddhist Era years (พ.ศ.), Thai numerals (๑๒๓), zero-width spaces that break tokenization, and mojibake from TIS-620 encoding. ThaiEDA catches all of these.

Privacy-first LLM analysis. Want to ask an LLM about your data but can't send raw rows to a cloud API? ThaiEDA has 4 privacy modes — the default sends zero raw data off your machine. Perfect for government, finance, and medical data under PDPA.

Insights, not just summaries. A cross-column insight engine finds non-obvious patterns — "column A strongly predicts column B", "this group is 3× higher than average" — ranked by statistical interestingness with Benjamini-Hochberg correction.

Thai-specific validation. National ID card checksum validation, Thai address parsing (province/district/subdistrict), Thai holiday awareness for timeseries spike attribution. No other EDA tool does this.

One line to get everything. thaieda.run(df) chains the full pipeline: type detection → smart cleaning → quality checks → anomaly detection → insight discovery → visualization → HTML report. No config needed.

---

How It Works

DataFrame
    │
    ▼
┌──────────────────────────────────────────────┐
│  thaieda.run(df)                             │
│                                              │
│  0. pre-analyze → data type + language       │
│  1. detect      → column types + Thai months │
│  2. clean       → smart cleaning (auto-decide)│
│  3. quality     → language-aware checks      │
│  4. anomaly     → statistical + ML + text    │
│  5. insights    → 6 cross-column patterns    │
│  6. viz         → interactive + static charts│
│  7. report      → executive HTML narrative   │

│  + optional: LLM analysis (4 privacy modes)  │
│  + optional: compare(df1, df2) side-by-side  │
└──────────────────────────────────────────────┘
    │
    ▼
EDAResult
  .to_html()        → report.html
  .to_dict()        → Python dict
  .to_json()        → JSON string
  .insights         → insight cards
  .cleaned_df       → cleaned DataFrame
  .quality_issues   → list of issues
  .quality_score    → 0-100 score with grade
  .anomalies        → anomaly findings
  .llm_response     → LLM analysis (if enabled)
  ._repr_html_()    → Jupyter rich display

run_folder("data/")  → FolderResult
  .to_html("dir/")      → individual HTML per file
  .to_master_html()     → single master HTML with sidebar
  .summary()            → text summary
  ._repr_html_()        → Jupyter rich display

---

What's New

Scale & Performance

Tested across 14 public datasets — from 500 rows to 541K rows, 8 to 171 columns. Every report stays under 2 MB and finishes under 120 seconds.

• Insight capping — reports surface the 30 most important findings instead of hundreds. Critical insights are always kept; warnings and info fill the rest. The executive summary shows the true count ("679 found, showing top 30").
• HTML bloat control — dual chart budget (40 charts max, 1.6 MB max). Quality and anomaly tables collapse after 50 rows. Wide tables switch to a summary view past 60 columns.
• Wide-table fast path — the insight engine samples breakdowns and measures when columns exceed 100. Correlation heatmaps and scatter matrices skip automatically on very wide data.
• Tall-table fast path — anomaly, quality, and outlier checks sample 50K rows when data exceeds 100K. Correlation computes on a sample. Timeseries decomposition skips past 200K rows.

Data Quality & Cleaning

• High-NA handling — columns over 80% missing are flagged as mostly_missing with NaN preserved. Columns over 40% get a warning to drop or impute with domain knowledge. Below 40% is unchanged.
• Smarter type detection — Thai low-cardinality text is classified as categorical, not free text. Text-named columns like review and feedback stay text even with few unique values.
• Cleaning safeguards — numeric strings like 1.00005 are left alone. Keyboard layout conversion only runs when Thai characters are present. Repeated-character spam on short codes is suppressed.
• ID/FK awareness — ID columns are excluded from categorical anomaly checks. *_id columns are detected even with low unique ratio. Buddhist Era checks skip IDs. Timeseries excludes ID/FK/code columns from measures.

Reporting

• Executive briefing format — reports flow from executive summary to key findings, business translation, priority actions, and plain-language explanations.
• Template pagination — Key Insights shows the top 20 with a collapsible section for the rest. Count badges are preserved.
• Fewer false positives — fuzzy duplicate guard skips short near-identical labels. Script mixing is skipped on low-cardinality columns. Outliers on heavy-tail distributions (skew > 2.0) are downgraded to info.
• Folder reports — run_folder() analyzes CSV, Excel, JSON, JSONL, and TSV folders. FolderResult.to_master_html() builds one master HTML with sidebar navigation.

Smart Pre-Analysis

• Language detection — Thai, English, mixed, and numeric data detected with confidence and per-column detail.
• Data type classification — transaction, registry, survey, timeseries, and mixed datasets classified before EDA.
• Language-aware quality — English-only data skips Thai-specific warnings automatically.

Accuracy Improvements (opt-in)

• Abbreviation expansion — expand_abbreviations() ขยายคำย่อไทย (กทม. → กรุงเทพมหานคร, บจ. → บริษัทจำกัด) ผ่าน pythainlp.util.abbreviation_to_full_text — opt-in operation ไม่ได้เปิดใช้ใน default pipeline เพราะเปลี่ยน semantics ของข้อความ
• Spell correction — spell_correct() แก้การสะกดคำผิดภาษาไทย (ขอบคุน → ขอบคุณ) ผ่าน pythainlp.spell.correct_sent — opt-in operation
• NFKC normalization — normalize_nfkc() แปลง full-width characters (Ａ→A, ９→9) ผ่าน stdlib unicodedata.normalize("NFKC") — opt-in operation
• Tokenizer selection modes — engine="auto-fast" เลือก nlpo3 (Rust, เร็ว 3-4x) และ engine="auto-quality" เลือก AttaCut (neural, แม่นยำสำหรับ social media/OOV text)
• Keyboard layout anomaly detection — ตรวจหาเซลล์ที่สงสัยว่าพิมพ์ผิด keyboard layout (ละตินผสมในคอลัมน์ไทย) — report-only ไม่แก้ไขอัตโนมัติ
• Thai grapheme validation — ตรวจหาวรรณยุกต์ซ้อนที่ผิดปกติ (เช่น ก่้ — mai ek + mai tho บนพยัญชนะเดียว) — report-only

---

Benchmarks

ThaiEDA is tested on 14 public datasets ranging from 500 rows to 541K rows, 8 to 171 columns. Every dataset produces a report under 2 MB in under 120 seconds.

Dataset	Rows	Cols	Time	HTML	Insights
titanic	891	12	8 s	0.79 MB	27
telco-churn	7,043	21	11 s	0.84 MB	11
wine-quality	1,599	12	7 s	0.93 MB	29
california-housing	20,640	10	15 s	0.99 MB	30
superstore	10,800	21	31 s	1.46 MB	30
adult	32,561	15	22 s	1.03 MB	29
bank-marketing	41,188	21	21 s	0.94 MB	30
online-retail	541,909	8	81 s	0.96 MB	30
dirty-thai-retail	500	8	2 s	0.51 MB	15
absenteeism	740	21	10 s	1.25 MB	30
online-shoppers	12,330	18	18 s	1.06 MB	30
aps-failure	16,000	171	100 s	0.48 MB	30
beijing-pm25	43,824	13	12 s	0.76 MB	19
bike-sharing	17,379	17	42 s	1.55 MB	30

All 14 datasets pass QA with 0 defects. Datasets from UCI ML Repository and public sources.

---

Examples

One-Line EDA

import thaieda
import pandas as pd

df = pd.read_csv("data.csv")

# Full pipeline in one call
result = thaieda.run(df)

# Access results
result.to_html("report.html")
print(result.quality_issues)
print(result.insights)

# In Jupyter: just display the result
result  # renders HTML report inline

Folder Mode — Analyze Every File at Once

import thaieda

# One line — analyzes every CSV/Excel/JSON in the folder
results = thaieda.run_folder("data/")

# Print summary
print(results.summary())
# ThaiEDA FolderResult — data/
#   Files: 5 (✅ 5 / ❌ 0)
#   ✅ customers.csv — 10,000 rows × 8 cols, 15 insights
#   ✅ orders.csv    — 50,000 rows × 12 cols, 28 insights
#   ...

# Save individual HTML reports
results.to_html("reports/")

# Generate a single master HTML with sidebar navigation
results.to_master_html("master-report.html")

run_folder() features:

• Auto-scans for CSV, Excel (.xlsx/.xls), JSON, JSONL, TSV
• recursive=True to include subfolders
• output_dir= to specify where HTML goes
• Error isolation — one broken file doesn't crash the rest
• progress= callback for progress tracking
• All run() kwargs supported (lang, clean, llm, etc.)
• to_master_html() — combines all reports into one page with sidebar nav + summary table

With LLM Analysis (Privacy-Safe)

import thaieda

# Default: zero raw data leaves your machine
result = thaieda.run(df, llm=True, privacy="insight_only", provider="ollama")
print(result.llm_response)

Compare Two Datasets

from thaieda.compare import compare_datasets

diff = compare_datasets(df_train, df_test, labels=("train", "test"))
print(diff["schema_diff"])      # columns added/removed
print(diff["drift"]["numeric"]) # KS statistic per column

Thai ID Card Validation

from thaieda.quality import validate_thai_id, validate_thai_id_column

# Single ID
validate_thai_id("1-1234-56789-01-2")  # → True/False

# Entire column
result = validate_thai_id_column(df["id_card"])
print(f"Valid: {result['valid_count']}, Invalid: {result['invalid_count']}")

Thai Address Parsing

from thaieda.detect import parse_thai_address

addr = parse_thai_address("123 หมู่ 4 ต.บางบัว อ.บางบัว จ.กรุงเทพฯ 10230")
print(addr)
# {'house_number': '123', 'moo': '4', 'subdistrict': 'บางบัว',
#  'district': 'บางบัว', 'province': 'กรุงเทพฯ', 'postal_code': '10230'}

Language Detection

import pandas as pd
from thaieda.detect import _detect_language

df = pd.DataFrame({
    "product": ["กาแฟ", "ชาไทย", "ขนม"],
    "review": ["อร่อยมาก 5/5 stars", "ดีครับ", "ไม่ดี"],
    "sku": ["SKU001", "SKU002", "SKU003"],
})

info = _detect_language(df)
print(info["language"], info["confidence"])
print(info["columns"])
# thai/mixed/english/numeric + per-column language map

Language Detection v2 features:

• Unicode Thai block analysis (U+0E00–U+0E7F) including vowels/tone marks (U+0E30–U+0E4D)
• Zero-width-space aware (\u200b, BOM, word joiner)
• Mixed-cell detection เช่น "อร่อยมาก 5/5 stars"
• Common Thai word hints: ครับ, ค่ะ, ไทย, อร่อย, ดี, ไม่, มี, และ
• Lazy pythainlp tokenizer when installed; regex fallback when unavailable
• Per-column column_details + dataset-level confidence (0.0–1.0)
• Sample-based scan (first 500 rows/column) for large DataFrames

Smart Pre-Analysis

ThaiEDA profiles the dataset before running the full report, so the narrative and quality checks match the data:

from thaieda.report import _detect_data_type

pre = _detect_data_type(df)
print(pre["label"], pre["language"]["language"])
print(pre["focus"])

Smart pre-analysis detects:

• Transaction data — orders, payments, revenue, invoices
• Registry/master data — customers, products, stores, entity attributes
• Survey/review data — ratings, comments, feedback text
• Timeseries data — datetime index/columns + numeric measures
• Mixed data — conservative fallback when signals overlap
• Language impact — Thai/mixed data enables Thai-specific checks; English-only data skips พ.ศ./เลขไทย checks automatically

Data Quality Score

from thaieda.quality import compute_quality_score

score = compute_quality_score(quality_issues, n_columns=10, n_rows=1000)
print(f"Score: {score.score}/100 ({score.grade})")
# Score: 85/100 (B)

Smart Cleaning

from thaieda.clean._smart import plan_cleaning

plan = plan_cleaning(df)
print(f"Actions: {plan.actions}")    # ['zwspace', 'numerals', 'duplicates']
print(f"Skipped: {plan.skipped}")    # ['encoding', 'whitespace']

---

Privacy Modes

Control exactly what data leaves your machine when using LLM analysis:

Mode	What Leaves	Guarantee	When to Use
insight_only (default)	Stats + insights only	Raw data never leaves	Government, medical, PDPA data
anonymized	Data with PII → tokens	Names/phones/ID cards masked	Need structure without raw PII
dp_noise	Stats + Laplace noise	Prevents re-identification	Small datasets where stats leak
full	Everything	None — you accept the risk	Public data, demos

---

What ThaiEDA Catches

Problem	Example	What Happens
Buddhist Era dates	15/03/2567	Auto-detects พ.ศ. → converts to CE
Thai numerals	๑๒๓ in numeric column	Converts to 123
Zero-width spaces	สม\u200bชาย	Strips invisible chars and reports language evidence
Thai vowel/tone marks	อร่อยค่ะ	Counts U+0E30–U+0E4D for better Thai detection
Mixed Thai/English cells	อร่อยมาก 5/5 stars	Detects as mixed language instead of English/numeric
Thai text in English-heavy tables	Thai product column + English IDs	Column-level language detection preserves Thai checks
Common Thai words	ครับ, ค่ะ, ไม่ดี	Boosts confidence for short Thai text
Mojibake encoding	à ¬Â¸Â¡Â¹	Auto-detects TIS-620 → UTF-8
Thai month names	มกราคม	Parses to ISO date
Phone numbers	081-234-5678	Detects + normalizes
National ID cards	1-1234-56789-01-2	Checksum validation
Thai addresses	123 ม.4 ต.บางบัว อ.บางบัว จ.กรุงเทพฯ	Parses to structured fields
Placeholder values	-, N/A, ไม่มี	Flags as missing
Constant columns	All same value	Flags as useless
Smart data type	Orders/reviews/timeseries	Pre-classifies transaction/registry/survey/timeseries/mixed
Language-aware checks	English-only DataFrame	Skips Thai-specific พ.ศ./เลขไทย warnings automatically
Thai holidays	Spike on Dec 5	Attributes to Father's Day
ID/FK semantics	order_id, store_id	Detected as ID even with low unique ratio; excluded from category anomaly
BE on ID	order_id=2531	No longer flagged as พ.ศ. (BE check requires date-like column)
Numeric string preservation	1.00005 in numeric data	fix_repeated_chars skips decimal/numeric strings
Keyboard layout guard	Floyd in English column	Not converted to Thai (requires Thai chars in column first)
Payment method detection	payment_method column	Classified as categorical, not amount column
Index artifact cleanup	Unnamed: 0 column	Ignored in analysis, flagged as index artifact
Per-column missing values	Age 20% missing	Quality issue with severity threshold (warning >5%, info 1-5%)
CSV delimiter warning	;-delimited file read as 1 column	Warns to re-read with sep=';'

---

Visualization

ThaiEDA generates both static (matplotlib) and interactive (Plotly) charts:

• Static: correlation heatmap, distribution, box/violin, missing matrix, scatter matrix, wordcloud, timeseries, pair plot, KDE, QQ plot, sunburst
• Interactive: hover tooltips, zoom, pan — using Plotly with Thai font (Sarabun) via Google Fonts
• Color palette: Okabe-Ito colorblind-safe (7 colors)
• Thai font: auto-detected for matplotlib, CSS-loaded for Plotly

from thaieda.viz._interactive import create_correlation_heatmap_interactive

html_div = create_correlation_heatmap_interactive(df)  # → HTML <div> for reports

---

Installation

# ติดตั้งทุกอย่างในคำสั่งเดียว
pip install thaieda

ไม่ต้องใส่ extras — pip install thaieda ติดตั้งทั้งหมด: Thai tokenizer, NER, ML, interactive charts, Excel, stats, encoding detection

LLM providers ยังเป็น optional (lazy-imported — ไม่ต้องติดตั้งถ้าไม่ใช้):

pip install openai       # OpenAI GPT
pip install anthropic    # Anthropic Claude
pip install ollama       # Ollama local LLM (หรือใช้ HTTP fallback ไม่ต้องติดตั้ง)

Requirements: Python 3.10+

---

Modules

Module	What It Does
run() / EDA()	One-liner API — full pipeline in one call
run_folder()	Analyze every CSV/Excel/JSON in a folder + master HTML
compare()	Side-by-side dataset comparison with drift detection
io/	Auto-read CSV/JSON/JSONL/Excel + encoding detection
detect/	Column type detection + Thai month names + address parsing + language detection v2
clean/	Smart cleaning: auto-decide what to fix (encoding, numerals, BE, zwspace)
quality/	Language-aware quality checks + score 0-100 + Thai ID card validation
anomaly/	Statistical + ML + text anomaly detection
ner/	Thai NER: person/place/organization
insight_engine/	6 cross-column insight patterns (BH-corrected)
viz/	Static + interactive charts with colorblind-safe palette
report/	Executive HTML report + smart pre-analysis (_detect_data_type)
llm/	Privacy-preserving LLM analysis (4 modes, 3 providers)
timeseries/	Trend/seasonality/STL/ACF + Thai holiday awareness
schema/	Multi-file PK/FK discovery + relationship matching

---

Testing

pytest tests/ -v                    # all tests (631 passed)
pytest tests/test_language_detection.py  # language detection + language-aware quality
pytest tests/test_thai_id.py        # ID card validation
pytest tests/test_thai_address.py   # address parsing
pytest tests/test_compare.py        # dataset comparison
pytest tests/test_run_folder.py     # folder mode + master HTML
pytest tests/test_llm.py            # LLM + privacy modes
ruff check src/ tests/              # lint
ruff format src/ tests/             # format

---

License

Apache-2.0 © Peet Wannasarnmetha