omni-eda 0.2.0

Fully automated exploratory data analysis for pandas DataFrames.

omni_eda

Fully automated, production-grade Exploratory Data Analysis for any pandas DataFrame.

omni_eda inspects a dataset, profiles every column, finds data-quality problems, computes statistics for every data type, generates dozens of visualizations, detects outliers and correlations, suggests feature engineering steps, optionally evaluates a target column, and renders everything into a polished, shareable HTML report — in one function call.

from omni_eda import OmniEDA

OmniEDA("data.csv").generate_report("report.html")

---

Contents

• Features
• Installation
• Quick start
• Command-line interface
• Configuration
• What's in the report
• Export formats
• Optional cleaning
• Architecture
• Performance notes
• Development
• License

---

Features

Feature / Capability	Description & Details
Broad Input Support	CSV/TSV, Excel, Parquet, Feather, JSON/JSONL, SQL (via SQLAlchemy connection), in-memory DataFrames, folders of files, or chunked CSV reader for datasets exceeding memory.
Automatic Column Profiling	Profiling for numeric, categorical, datetime, boolean, text, and ID columns. Semantic detection of emails, URLs, phone numbers, currencies, percentages, coordinates, ZIP codes, country/state/city columns, binary/label/ordinal-encoded columns, constant/high-cardinality columns, and mixed-dtype columns.
Data Quality Report	Detection of missing values, duplicate rows/columns, infinities, impossible negatives, invalid/future dates, empty/whitespace-only strings, hidden/non-printable characters, encoding issues, skewed distributions, class imbalance, highly correlated columns, and target-leakage candidates.
Descriptive Statistics	Mean/median/mode/std/MAD/IQR/percentiles/skewness/kurtosis for numeric columns; frequency/entropy/rare-category detection for categoricals; length/word/character stats for free text; range/seasonality for datetimes.
~40 Plot Types	Univariate, bivariate, multivariate, and time-series analysis (histograms/KDE, boxplots, violin, ECDF, Q-Q, rank, lollipop, pie, scatter, hexbin, regression, residual, joint, pairplots, mosaic, cross-tab heatmaps, FacetGrid, cluster maps, parallel coordinates, Andrews curves, radar, bubble, 3D scatter, PCA/t-SNE/UMAP, correlation networks, trend, seasonality, rolling mean/std, lag plots, ACF/PACF, etc.). Plotting is skipped automatically when it wouldn't make sense.
Correlation & Association	Pearson/Spearman/Kendall, Cramer's V, correlation ratio (categorical ↔ numeric), mutual information, and distance correlation.
Outlier Detection	Z-score, modified Z-score, IQR, Isolation Forest, Local Outlier Factor, Elliptic Envelope, and DBSCAN.
Feature Engineering Suggestions	Encoding, scaling, log/power transforms, binning, datetime decomposition, interaction & polynomial feature candidates, redundant-feature and rare-category flags.
Target Analysis	Class imbalance, ANOVA / chi-square association tests, Random-Forest feature importance + mutual information, and (for binary targets) a baseline classifier with ROC, Precision-Recall, and lift charts.
Multi-Format Export	HTML report, Markdown, JSON, Excel workbook, PDF (figure bundle), raw PNG/SVG figures, CSV tables, and a self-contained interactive HTML dashboard.
Built for Scale	Sampling guards, vectorized pandas/NumPy operations, optional multiprocessing, memory-aware dtype downcasting, and defensive handling of empty, single-row, single-column, and all-null-column datasets.
Optional, Auditable Cleaning	Every cleaning step is opt-in, logged, and returns a new DataFrame; nothing is changed silently.

Installation

pip install omni-eda

or, from a local checkout:

pip install -e .

Some visualizations (mosaic plots, ACF/PACF, correlation networks, UMAP) and file formats (Parquet/Feather) need extra libraries. Install everything with:

pip install "omni-eda[extra]"

Every feature that needs an extra dependency degrades gracefully (it's skipped, with a log message) if that dependency isn't installed — nothing crashes.

Quick start

import pandas as pd
from omni_eda import OmniEDA

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

eda = OmniEDA()
results = eda.run(df)          # run the full pipeline
eda.summary()                  # quick console overview
eda.generate_report("report.html")

One-liner form:

from omni_eda import OmniEDA

OmniEDA("customers.csv").generate_report("report.html")

With a target column (enables class-imbalance checks, feature importance, ANOVA/chi-square tests, and ROC/PR/lift curves for binary targets):

from omni_eda import OmniEDA, EDAConfig

config = EDAConfig(target_column="churned", theme="corporate")
eda = OmniEDA(config=config)
eda.run(df)
eda.export(formats=["html", "excel", "dashboard"])

See examples/basic_usage.py and examples/advanced_usage.py for complete, runnable scripts.

Command-line interface

# Analyze a file and write an HTML report
omni-eda run data.csv -o report_output -f html

# With a target column, several export formats, and a lighter run for a quick look
omni-eda run data.csv --target price -f html json excel --sample-rows 50000

# Run just the (conservative) auto-cleaning pipeline
omni-eda clean data.csv -o cleaned.csv

Run omni-eda run --help for the full list of flags (theme, ignored columns, output formats, quiet mode, etc.).

Configuration

Every tunable knob lives in a single EDAConfig dataclass (see omni_eda/config.py for the full list with defaults):

from omni_eda import EDAConfig

config = EDAConfig(
    title="Customer Churn - EDA",
    target_column="churned",
    ignore_columns=["customer_id"],
    theme="dark",                          # light | dark | minimal | corporate
    high_correlation_threshold=0.85,
    outlier_methods=["iqr", "zscore", "isolation_forest", "lof"],
    sample_for_plots=20_000,               # cap rows used for plotting on huge datasets
    max_rows_for_expensive_ops=50_000,     # cap rows used for correlation/outlier/model fits
    export_formats=["html", "json", "excel"],
)

What's in the report

The generated HTML report includes: a dataset overview (rows, columns, memory, column-type breakdown), the full data-quality issue list (severity-tagged), missing-value analysis with matrix/heatmap/dendrogram/bar visualizations, per-column statistics and distribution plots, correlation heatmaps and high-correlation pairs, an outlier summary table, bivariate and multivariate visualization galleries, time-series analysis (when datetime columns are present), target analysis (when a target column is configured), feature-engineering suggestions, and a plain-language final summary.

Export formats

Format	What you get
html	The full visual report (self-contained, images embedded as base64)
markdown	A text-only version of the report (tables + suggestions, no images)
json	Machine-readable statistics, quality issues, correlations, suggestions
excel	A multi-sheet workbook (overview, statistics, issues, correlations, ...)
pdf	Every generated PNG figure bundled into a single multi-page PDF
csv	Separate CSVs for statistics, quality issues, correlations, outliers
figures	Every individual plot as a standalone PNG/SVG file
dashboard	A single-file interactive HTML dashboard (Plotly.js via CDN, no server)

eda.export(output_dir="out", formats=["html", "excel", "dashboard"])

Optional cleaning

Cleaning never happens implicitly. Call .clean() explicitly and get back a new DataFrame plus a human-readable log of exactly what changed:

cleaned_df = eda.clean(steps=["dedup_rows", "dedup_columns", "convert_dtypes", "infinities"])

Available steps: dedup_rows, dedup_columns, drop_constant, fill_missing, convert_dtypes, strip_whitespace, infinities (the default pipeline runs the non-destructive subset of these; fill_missing and drop_constant are opt-in since they change the data more aggressively). Lower-level building blocks (clip_outliers_iqr, encode_categoricals, scale_numeric_columns, ...) are available directly from omni_eda.cleaning for custom pipelines.

Architecture

omni_eda/
├── __init__.py            # public API: OmniEDA, EDAConfig
├── analyzer.py             # OmniEDA orchestrator - runs the full pipeline
├── config.py                # EDAConfig dataclass (every tunable setting)
├── loaders.py                # CSV/Excel/Parquet/Feather/JSON/SQL/folder loading
├── detection.py               # column type & semantic-role detection
├── statistics.py               # descriptive statistics per dtype
├── cleaning.py                   # optional, auditable cleaning operations
├── quality.py                     # data quality issue detection
├── correlation.py                  # Pearson/Spearman/Kendall/Cramer's V/MI/distance corr
├── outliers.py                      # Z-score/IQR/Isolation Forest/LOF/DBSCAN/Elliptic Env.
├── missing.py                        # missing-value analysis & visualization
├── visualization.py                   # ~40 plot functions + the PlotEngine orchestrator
├── feature_engineering.py              # rule-based feature suggestions
├── target_analysis.py                   # class imbalance, tests, importance, ROC/PR/lift
├── themes.py                             # matplotlib/seaborn/report color themes
├── report.py                              # HTML/Markdown/console report builder (Jinja2)
├── export.py                               # HTML/MD/JSON/Excel/PDF/CSV/figure export
├── dashboard.py                             # self-contained interactive HTML dashboard
├── logger.py                                 # package-wide logging + progress bars
├── utils.py                                   # shared helpers (sampling, caching, ...)
├── cli.py                                      # `omni-eda` command-line interface
└── templates/report.html.j2                     # the HTML report template

Every pipeline stage in OmniEDA.run() is wrapped so a failure in one stage (an exotic dtype breaking a single plot, say) is logged and skipped rather than aborting the whole run.

Performance notes

omni_eda is built to stay usable on large datasets without needing a distributed runtime:

• Expensive operations (correlation, outlier detection, PCA/t-SNE, model fitting) sample down to max_rows_for_expensive_ops (default 50,000 rows) rather than processing everything.
• Plotting samples down to sample_for_plots (default 20,000 rows).
• Duplicate-column detection hashes columns instead of transposing the DataFrame, which is dramatically faster on wide-ish, long DataFrames.
• CSV loading supports chunksize for files too large to read at once.
• Numeric downcasting and category-dtype conversion (omni_eda.utils.optimize_dtypes) are available to cut memory usage before analysis.
• Random Forest / mutual information calls in target analysis use their own bounded samples so a 26-column, 100k-row dataset with a target column finishes in roughly a minute on a single CPU core; tune max_rows_for_expensive_ops, n_jobs, and enable_target_modeling / enable_model_based_outliers down further for a quicker look at very large data.

Development

git clone https://github.com/example/omni_eda.git
cd omni_eda
pip install -e ".[dev,extra]"
pre-commit install

pytest                       # run the test suite
pytest --cov=omni_eda        # with coverage
ruff check omni_eda tests    # lint
black omni_eda tests         # format
mypy omni_eda                # type-check

Contributions are welcome — please open an issue or pull request.

License

MIT — see LICENSE.