transfory 1.3.0

An Object-Oriented Data Transformation Toolkit

Transfory

An Object-Oriented, Explainable Data Transformation Toolkit for Python.

Transfory is a data preprocessing library designed with clarity and modularity in mind. Inspired by scikit-learn's API, it provides a suite of common data transformation tools that can be easily combined into powerful, reusable pipelines.

What makes Transfory unique is its built-in InsightReporter, an explainability engine that provides detailed, human-readable logs of every step in your transformation process. No more black boxes—understand exactly what happens to your data.

Key Features

• Modular Transformers: A collection of intuitive, single-purpose transformers for common preprocessing tasks.
• Powerful Pipelines: Chain transformers together with Pipeline and apply different logic to different columns with ColumnTransformer.
• Explainability First: The InsightReporter gives you a step-by-step narrative of your data's journey, making debugging and validation effortless.
• Pandas-Native: Built to work seamlessly with pandas DataFrames.

Installation

You can install Transfory directly from PyPI:

pip install transfory

Quick Start

Let's perform a simple data cleaning operation on the Titanic dataset and see what Transfory tells us.

import pandas as pd
import seaborn as sns
from transfory.pipeline import Pipeline
from transfory.missing import MissingValueHandler
from transfory.encoder import Encoder
from transfory.scaler import Scaler
from transfory.insight import InsightReporter

# 1. Load data and initialize the reporter
df = sns.load_dataset('titanic')[['age', 'fare', 'embarked', 'sex']]
reporter = InsightReporter()

# 2. Define a multi-step pipeline
pipeline = Pipeline(
    steps=[
        ("imputer", MissingValueHandler(strategy="mean")),
        ("encoder", Encoder(method="onehot")),
        ("scaler", Scaler(method="zscore"))
    ],
    logging_callback=reporter.get_callback()
)

# 3. Fit and transform the data
transformed_df = pipeline.fit_transform(df)

print("Transformed Data:")
print(transformed_df.head())

# 4. Review the Insight Report for a step-by-step explanation
print("\n--- Insight Report ---")
print(reporter.summary())

The InsightReporter will output a clear summary of what happened, such as the value used for imputation, the new columns created by the encoder, and the columns that were scaled.

Core Modules Explained

Transfory is built on a set of core components that can be mixed and matched to create any preprocessing workflow.

Orchestrators

These modules are used to control the flow and application of transformers.

Pipeline

Sequentially applies a list of transformers. The output of one step becomes the input to the next, making it easy to define an ordered workflow.

from transfory.pipeline import Pipeline

pipeline = Pipeline([
    ("imputer", MissingValueHandler()),
    ("scaler", Scaler()),
])

ColumnTransformer

Applies different transformers to different columns of a DataFrame in parallel. This is highly efficient and essential for workflows where numeric and categorical data need separate handling.

from transfory.column_transformer import ColumnTransformer

preprocessor = ColumnTransformer(
    transformers=[
        ("numeric_scaler", Scaler(), ['age', 'fare']),
        ("categorical_encoder", Encoder(), ['embarked', 'sex'])
    ],
    remainder='passthrough' # Keep other columns
)

InsightReporter

The explainability engine of Transfory. It captures events from all transformers in a pipeline and provides a detailed, human-readable summary of the entire process. Simply create an instance and pass its callback to your pipeline.

from transfory.insight import InsightReporter

reporter = InsightReporter()
pipeline = Pipeline([...], logging_callback=reporter.get_callback())
pipeline.fit_transform(df)
print(reporter.summary())

---

Data Transformers

These are the building blocks that perform the actual data transformations.

MissingValueHandler

Handles missing (NaN) values in a DataFrame.

• Strategies: mean, median, mode, constant (with fill_value).

import pandas as pd
import numpy as np
from transfory.missing import MissingValueHandler

data = pd.DataFrame({'score': [88, 92, np.nan, 75], 'grade': ['A', 'A', 'C', np.nan]})

# Impute missing 'score' with the mean and 'grade' with the mode
imputer = MissingValueHandler(strategy="mean")
transformed_data = imputer.fit_transform(data)
print(transformed_data)

Encoder

Converts categorical columns into a numerical format.

• Methods: onehot (creates binary columns), label (assigns a unique integer to each category).

import pandas as pd
from transfory.encoder import Encoder

data = pd.DataFrame({'color': ['red', 'blue', 'green', 'red']})

# Convert the 'color' column into one-hot encoded format
encoder = Encoder(method="onehot")
transformed_data = encoder.fit_transform(data)
print(transformed_data)

Scaler

Scales numeric features to a common range.

• Methods: zscore (standard scaling), minmax (scales to a [0, 1] range).

from transfory.scaler import Scaler

# Scale features to have zero mean and unit variance
scaler = Scaler(method="zscore")

OutlierHandler

Caps extreme values (outliers) to mitigate their effect on downstream models.

• Methods: iqr (caps at 1.5 * Interquartile Range), zscore (caps at a z-score threshold), quantile.

from transfory.outlier import OutlierHandler

# Cap values outside the 1.5 * IQR range
outlier_handler = OutlierHandler(method="iqr")

DatetimeFeatureExtractor

Creates new features from datetime columns, such as year, month, day, or day of the week.

from transfory.datetime import DatetimeFeatureExtractor

# Extract year and month from a 'transaction_date' column
date_extractor = DatetimeFeatureExtractor(features=['year', 'month'])

FeatureGenerator

Automatically creates polynomial and interaction features from numeric columns to help models capture non-linear relationships.

from transfory.featuregen import FeatureGenerator

# Create quadratic features (e.g., a^2, b^2) and interaction features (a*b)
feature_gen = FeatureGenerator(degree=2, include_interactions=True)

Comparison: Scikit-Learn vs Transfory

Transfory is inspired by scikit-learn, but it is built specifically to be more transparent, Pandas-native, and beginner-friendly. This table summarizes the key differences.

Feature Comparison Table

Aspect	Scikit-Learn	Transfory
Design Philosophy	Highly optimized ML preprocessing and modeling framework.	Explainable, modular, beginner-friendly data transformation toolkit.
Data Structure Support	Primarily works with NumPy arrays. Pandas support is secondary.	100% Pandas-native. Always returns DataFrames with column names preserved.
Explainability	Minimal. Transformers do not log what they did unless manually wrapped.	Built-in InsightReporter logs every step in human-readable form.
Pipeline System	Powerful, but verbose and produces NumPy arrays by default.	Simpler, cleaner, DataFrame-preserving pipelines.
ColumnTransformer	Flexible but difficult to inspect; output column names often tricky.	Column names preserved, easy to debug, supports explainability.
Error Messages	Often cryptic (NumPy errors, shape mismatch errors).	Human-friendly, transformer-specific error messages.
Extensibility	Requires subclassing and adhering to strict API.	Extremely easy—just subclass BaseTransformer.
Transform Output	NumPy array (unless wrapped).	Pandas DataFrame (always).
Logging	None by default.	Automatic logging through callbacks.
Intended Users	Production ML engineers, researchers.	Students, data analysts, instructors, prototyping ML pipelines.

Demos

For more in-depth examples, check out the Jupyter Notebooks in the /demo directory. They provide interactive, step-by-step walkthroughs of the library's capabilities on real datasets.

1. demo1_basic_pipeline.ipynb: Introduction to Pipeline.
2. demo2_column_transformer.ipynb: Advanced workflows with ColumnTransformer.
3. demo3_feature_engineering.ipynb: Using DatetimeFeatureExtractor and FeatureGenerator.
4. demo4_outlier_handling.ipynb: Using the OutlierHandler.

Contributing

Contributions are welcome! If you have a feature request, bug report, or want to improve the documentation, please open an issue or submit a pull request on the GitHub repository.

Contributors

This project was made possible by the following contributors:

Project Lead & Core Architecture

• Rogelio Q. Mandamian III ([e-mail])
  • Project lead, architect of the core framework (BaseTransformer, Pipeline, ColumnTransformer), and developer of the InsightReporter.

Module Development

• Jose Gabriel P. Yana ([e-mail])
  • Responsible for the implementation of Scaler, MissingValueHandler, Encoder, and FeatureGenerator.
• Renell josh cuerquis ([e-mail])
  • Responsible for the implementation of Scaler, MissingValueHandler, Encoder, and FeatureGenerator.

Documentation

• Ashlanie L. Pandi ([e-mail])
  • Responsible for the creation and refinement of all user-facing documentation.
• Eleah Jaizel T. Cabili ([e-mail])
  • Responsible for the creation and refinement of all user-facing documentation.

License

This project is licensed under the MIT License. See the LICENSE file for details.