DShandler 0.1.2

A stateful, chainable data science pipeline for tabular ML workflows.

dspipeline

A stateful, chainable data-science pipeline for the full tabular ML workflow — from raw data to model-ready arrays — in a single class.

Installation

pip install dspipeline

Quick start

import pandas as pd
from dspipeline import DataSciencePipeline

df  = pd.read_csv("your_dataset.csv")
dsp = DataSciencePipeline(df, target_col="Churn", task_type="classification")

# One-liner: diagnostics → cleaning → preprocessing
dsp.run_diagnostics().run_cleaning().run_preprocessing()

# Leakproof split
X_train, X_test, y_train, y_test, preprocessor = dsp.split(test_size=0.2)

# Fit your model on processed arrays
from sklearn.ensemble import GradientBoostingClassifier
model = GradientBoostingClassifier()
model.fit(dsp.results["split"]["X_train_processed"], y_train)

---

What it covers

Phase	Methods
Diagnostics	profile_missing, detect_structural, detect_dimensional, detect_categorical, detect_predictive, detect_anomaly_scan, detect_leakage
Cleaning	format_structure, drop_duplicates, standardize_text, impute_numeric, impute_categorical
Anomaly handling	handle_outliers
Transformation	transform_shape
Encoding	encode
Feature selection	select_features, vif_optimize
Split	split
EDA	analyze_distribution, evaluate_distribution, analyze_relationship, test_hypothesis
Time-series	enforce_stationarity, analyze_autocorrelation

---

Method chaining

Every mutating method returns self, so you can chain calls:

(dsp
  .profile_missing()
  .drop_duplicates(subset=["user_id"], sort_by="updated_at")
  .impute_numeric(strategy="knn")
  .impute_categorical(strategy="mode")
  .handle_outliers(method="iqr", action="clip", threshold=1.5)
  .transform_shape(scale_method="robust")
  .encode(nominal_cols=["color"], ordinal_maps={"size": ["S", "M", "L"]})
  .select_features(multi_corr_threshold=0.85)
  .vif_optimize(threshold=5.0)
)
X_train, X_test, y_train, y_test, preprocessor = dsp.split(stratify=True)

---

State inspection

dsp.summary()           # prints shape history for every step
dsp.results.keys()      # all stored reports and artefacts
dsp.history             # list of {method, shape_before, shape_after}
df_snapshot = dsp.snapshot()   # copy of current working DataFrame
dsp.reset()             # restore to original raw DataFrame

---

Individual functions

Every function is also importable directly:

from dspipeline import (
    advanced_missing_profiler,
    detect_anomalies,
    handle_numerical_missing,
    advanced_knn_impute,
    transform_data_shape,
    encode_categorical_data,
    optimize_vif,
    setup_leakproof_environment,
    analyze_distribution,
    test_hypothesis,
    enforce_stationarity,
    analyze_autocorrelation,
)

---

Time-series example

dsp = DataSciencePipeline(df, target_col="demand")

# Check and fix stationarity
series, d, report = dsp.enforce_stationarity("revenue", seasonal_period=12)
print(f"Applied d={d} differencing steps")

# ACF / PACF + ARIMA order hints
acf_report = dsp.analyze_autocorrelation("revenue", lags=40)
print(f"Suggested ARIMA({acf_report['arima_hint_p']}, {d}, {acf_report['arima_hint_q']})")

---

Requirements

• Python ≥ 3.9
• pandas, numpy, scikit-learn, scipy, statsmodels, matplotlib, seaborn

---

License

MIT