Easily find and add relevant features to your ML & AI pipeline from
hundreds of public, community, and premium external data sources,
including open & commercial LLMs

Quick Start in Colab » | Register / Sign In | Slack Community | Propose a new data source

Upgini is an intelligent data search engine with a Python library that helps you find and add relevant features to your ML pipeline from hundreds of public, community, and premium external data sources. Under the hood, Upgini automatically optimizes all connected data sources by generating an optimal set of ML features using large language models (LLMs), GNNs (graph neural networks), and recurrent neural networks (RNNs).

Motivation: for most supervised ML models external data & features boost accuracy significantly better than any hyperparameters tuning. But lack of automated and time-efficient enrichment tools for external data blocks massive adoption of external features in ML pipelines. We want to radically simplify feature search and enrichment to make external data a standard approach. Like hyperparameter tuning in machine learning today.

Mission: Democratize access to data sources for data science community.

⭐️ Automatically find only relevant features that improve your model’s accuracy. Not just correlated with the target variable, which in 9 out of 10 cases yields zero accuracy improvement
⭐️ Automated feature generation from the sources: feature generation with LLM‑based data augmentation, RNNs, and GraphNNs; ensembling across multiple data sources
⭐️ Automatic search key augmentation from all connected sources. If you do not have all search keys in your search request, such as postal/ZIP code, Upgini will try to add those keys based on the provided set of search keys. This will broaden the search across all available data sources
⭐️ Calculate accuracy metrics and uplift after enriching an existing ML model with external features
⭐️ Check the stability of accuracy gain from external data on out-of-time intervals and verification datasets. Mitigate the risks of unstable external data dependencies in the ML pipeline
⭐️ Easy to use - a single request to enrich the training dataset with all of the keys at once:

⭐️ Scikit-learn-compatible interface for quick data integration with existing ML pipelines
⭐️ Support for most common supervised ML tasks on tabular data:

⭐️ Simple Drag & Drop Search UI:

• Public data: public sector, academic institutions, other sources through open data portals. Curated and updated by the Upgini team
• Community‑shared data: royalty- or license-free datasets or features from the data science community (our users). This includes both public and scraped data
• Premium data providers: commercial data sources verified by the Upgini team in real-world use cases

👉 Details on datasets and features

❓Know other useful data sources for machine learning? Give us a hint and we'll add it for free.

• The goal is to predict salary for a data science job posting based on information about the employer and job description.
• Following this guide, you'll learn how to search and auto‑generate new relevant features with the Upgini library
• The evaluation metric is Mean Absolute Error (MAE).

Run Feature search & generation notebook inside your browser:

• The goal is to predict future sales of different goods in stores based on a 5-year history of sales.
• Kaggle Competition Store Item Demand Forecasting Challenge is a product sales forecasting competition. The evaluation metric is SMAPE.

Run Simple sales prediction for retail stores inside your browser:

• The goal is to improve a Top‑1 winning Kaggle solution by adding new relevant external features and data.
• Kaggle Competition is a product sales forecasting competition; the evaluation metric is SMAPE.

• Save time on feature search and engineering. Use ready-to-use external features and data sources to maximize overall AutoML accuracy, right out of the box.
• Kaggle Competition is a product sales forecasting, evaluation metric is SMAPE.
• Low-code AutoML frameworks: Upgini and PyCaret

• The goal is to improve the accuracy of multivariate time‑series forecasting using new relevant external features and data. The main challenge is the data and feature enrichment strategy, in which a component of a multivariate time series depends not only on its past values but also on other components.
• Kaggle Competition is a product sales forecasting, evaluation metric is RMSLE.

• Save time on external data wrangling and feature calculation code for hypothesis tests. The key challenge is the time‑dependent representation of information in the training dataset, which is uncommon for credit default prediction tasks. As a result, special data enrichment strategy is used.
• Kaggle Competition is a credit default prediction, evaluation metric is normalized Gini coefficient.

%pip install upgini

You can use your labeled training datasets "as is" to initiate the search. Under the hood, we'll search for relevant data using:

• search keys from the training dataset to match records from potential data sources with new features
• labels from the training dataset to estimate the relevance of features or datasets for your ML task and calculate feature importance metrics
• your features from the training dataset to find external datasets and features that improve accuracy of your existing data and estimate accuracy uplift (optional)

Load the training dataset into a Pandas DataFrame and separate feature columns from the label column in a Scikit-learn way:

import pandas as pd
# labeled training dataset - customer_churn_prediction_train.csv
train_df = pd.read_csv("customer_churn_prediction_train.csv")
X = train_df.drop(columns="churn_flag")
y = train_df["churn_flag"]

Search keys columns will be used to match records from all potential external data sources/features.
Define one or more columns as search keys when initializing the FeaturesEnricher class.

from upgini.features_enricher import FeaturesEnricher
from upgini.metadata import SearchKey

enricher = FeaturesEnricher(
	search_keys={
		"subscription_activation_date": SearchKey.DATE,
		"country": SearchKey.COUNTRY,
		"zip_code": SearchKey.POSTAL_CODE,
		"hashed_email": SearchKey.HEM,
		"last_visit_ip_address": SearchKey.IP,
		"registered_with_phone": SearchKey.PHONE
	})

For the search key types SearchKey.DATE/SearchKey.DATETIME with dtypes object or string you have to specify the date/datetime format by passing date_format parameter to FeaturesEnricher. For example:

from upgini.features_enricher import FeaturesEnricher
from upgini.metadata import SearchKey

enricher = FeaturesEnricher(
	search_keys={
		"subscription_activation_date": SearchKey.DATE,
		"country": SearchKey.COUNTRY,
		"zip_code": SearchKey.POSTAL_CODE,
		"hashed_email": SearchKey.HEM,
		"last_visit_ip_address": SearchKey.IP,
		"registered_with_phone": SearchKey.PHONE
	}, 
	date_format = "%Y-%d-%m"
)

To use a non-UTC timezone for datetime, you can cast datetime column explicitly to your timezone (example for Warsaw):

df["date"] = df.date.astype("datetime64").dt.tz_localize("Europe/Warsaw")

A single country for the whole training dataset can be passed via country_code parameter:

from upgini.features_enricher import FeaturesEnricher
from upgini.metadata import SearchKey

enricher = FeaturesEnricher(
	search_keys={
		"subscription_activation_date": SearchKey.DATE,
		"zip_code": SearchKey.POSTAL_CODE,
	}, 
	country_code = "US",
	date_format = "%Y-%d-%m"
)

The main abstraction you interact with is FeaturesEnricher, a Scikit-learn-compatible estimator. You can easily add it to your existing ML pipelines. Create an instance of the FeaturesEnricher class and call:

• fit to search relevant datasets & features
• then transform to enrich your dataset with features from the search result

Let's try it out!

import pandas as pd
from upgini.features_enricher import FeaturesEnricher
from upgini.metadata import SearchKey

# load labeled training dataset to initiate search
train_df = pd.read_csv("customer_churn_prediction_train.csv")
X = train_df.drop(columns="churn_flag")
y = train_df["churn_flag"]

# now we're going to create an instance of the `FeaturesEnricher` class
enricher = FeaturesEnricher(
	search_keys={
		"subscription_activation_date": SearchKey.DATE,
		"country": SearchKey.COUNTRY,
		"zip_code": SearchKey.POSTAL_CODE
	})

# Everything is ready to fit! For 100k records, fitting should take around 10 minutes
# We'll send an email notification; just register on profile.upgini.com
enricher.fit(X, y)

That's it! The FeaturesEnricher is now fitted.

FeaturesEnricher class has two properties for feature importances, that are populated after fit - feature_names_ and feature_importances_:

• feature_names_ - feature names from the search result, and if parameter keep_input=True was used, initial columns from search dataset as well
• feature_importances_ - SHAP values for features from the search result, same order as in feature_names_

Method get_features_info() returns pandas dataframe with features and full statistics after fit, including SHAP values and match rates:

enricher.get_features_info()

Get more details about FeaturesEnricher at runtime using docstrings via help(FeaturesEnricher) or help(FeaturesEnricher.fit).

FeaturesEnricher is a Scikit-learn-compatible estimator, so any pandas dataframe can be enriched with external features from a search result (after fit).
Use the transform method of FeaturesEnricher, and let the magic do the rest 🪄

# load dataset for enrichment
test_x = pd.read_csv("test.csv")
# enrich it!
enriched_test_features = enricher.transform(test_x)

FeaturesEnricher can be initialized with search_id from a completed search (after a fit call). Just use enricher.get_search_id() or copy search id string from the fit() output.
Search keys and features in X must be the same as for fit()

enricher = FeaturesEnricher(
  # same set of search keys as for the fit step
  search_keys={"date": SearchKey.DATE},
  api_key="<YOUR API_KEY>",  # if you fitted the enricher with an api_key, then you should use it here
  search_id = "abcdef00-0000-0000-0000-999999999999"
)
enriched_prod_dataframe = enricher.transform(input_dataframe)

In most ML cases, the training step requires a labeled dataset with historical observations. For production, you'll need updated, current data sources and features to generate predictions.
FeaturesEnricher, when initialized with a set of search keys that includes SearchKey.DATE, will match records from all potential external data sources exactly on the specified date/datetime based on SearchKey.DATE, to avoid enrichment with features "from the future" during the fit step.
And then, for transform in a production ML pipeline, you'll get enrichment with relevant features, current as of the present date.

⚠️ Include SearchKey.DATE in the set of search keys to get current features for production and avoid features from the future during training:

enricher = FeaturesEnricher(
	search_keys={
		"subscription_activation_date": SearchKey.DATE,
		"country": SearchKey.COUNTRY,
		"zip_code": SearchKey.POSTAL_CODE,
	},
) 

We validate and clean the search‑initialization dataset under the hood:

• check your search keys columns' formats;
• check zero variance for label column;
• check dataset for full row duplicates. If we find any, we remove them and report their share;
• check inconsistent labels - rows with the same features and keys but different labels, we remove them and report their share;
• remove columns with zero variance - we treat any non search key column in the search dataset as a feature, so columns with zero variance will be removed

We detect ML task under the hood based on label column values. Currently we support:

• ModelTaskType.BINARY
• ModelTaskType.MULTICLASS
• ModelTaskType.REGRESSION

But for certain search datasets you can pass parameter to FeaturesEnricher with correct ML task type:

from upgini.features_enricher import FeaturesEnricher
from upgini.metadata import SearchKey, ModelTaskType

enricher = FeaturesEnricher(
	search_keys={"subscription_activation_date": SearchKey.DATE},
	model_task_type=ModelTaskType.REGRESSION
)

Time-series prediction is supported as ModelTaskType.REGRESSION or ModelTaskType.BINARY tasks with time-series‑specific cross-validation splits:

• Scikit-learn time-series cross-validation - CVType.time_series parameter
• Blocked time-series cross-validation - CVType.blocked_time_series parameter

To initiate feature search, you can pass the cross-validation type parameter to FeaturesEnricher with a time-series‑specific CV type:

from upgini.features_enricher import FeaturesEnricher
from upgini.metadata import SearchKey, CVType

enricher = FeaturesEnricher(
	search_keys={"sales_date": SearchKey.DATE},
	cv=CVType.time_series
)

If you're working with multivariate time series, you should specify id columns of individual univariate series in FeaturesEnricher. For example, if you have a dataset predicting sales for different stores and products, you should specify store and product id columns as follows:

enricher = FeaturesEnricher(
		search_keys={
				"sales_date": SearchKey.DATE,
    },
    id_columns=["store_id", "product_id"],
    cv=CVType.time_series
)

⚠️ Preprocess the dataset in case of time-series prediction:
sort rows in dataset according to observation order, in most cases - ascending order by date/datetime.

FeaturesEnricher automatically calculates model metrics and uplift from new relevant features either using calculate_metrics() method or calculate_metrics=True parameter in fit or fit_transform methods (example below).
You can use any model estimator with scikit-learn-compatible interface, some examples are:

• All Scikit-Learn supervised models
• Xgboost
• LightGBM
• CatBoost

In addition to that list, you can define a custom evaluation metric function using scikit-learn make_scorer, for example SMAPE.

By default, the calculate_metrics() method calculates the evaluation metric with the same cross-validation split as selected for FeaturesEnricher.fit() by the parameter cv = CVType.<cross-validation-split>.
But you can easily define a new split by passing a subclass of BaseCrossValidator to the cv parameter in calculate_metrics().

Example with more tips-and-tricks:

from upgini.features_enricher import FeaturesEnricher
from upgini.metadata import SearchKey

enricher = FeaturesEnricher(search_keys={"registration_date": SearchKey.DATE})

# Fit with default setup for metrics calculation
# CatBoost will be used
enricher.fit(X, y, eval_set=eval_set, calculate_metrics=True)

# LightGBM estimator for metrics
custom_estimator = LGBMRegressor()
enricher.calculate_metrics(estimator=custom_estimator)

# Custom metric function to scoring param (callable or name)
custom_scoring = "RMSLE"
enricher.calculate_metrics(scoring=custom_scoring)

# Custom cross validator
custom_cv = TimeSeriesSplit(n_splits=5)
enricher.calculate_metrics(cv=custom_cv)

# All of these custom parameters can be combined in both methods: fit, fit_transform and calculate_metrics:
enricher.fit(X, y, eval_set, calculate_metrics=True, estimator=custom_estimator, scoring=custom_scoring, cv=custom_cv)

If a training dataset has a text column, you can generate additional embeddings from it using instruction‑guided embedding generation with LLMs and data augmentation from external sources, just like Upgini does for all records from connected data sources.

In most cases, this gives better results than direct embeddings generation from a text field. Currently, Upgini has two LLMs connected to the search engine - GPT-3.5 from OpenAI and GPT-J.

To use this feature, pass the column names as arguments to the generate_features parameter. You can use up to 2 columns.

Here's an example for generating features from the "description" and "summary" columns:

enricher = FeaturesEnricher(
    search_keys={"date": SearchKey.DATE},
    generate_features=["description", "summary"]
)

With this code, Upgini will generate LLM embeddings from text columns and then check them for predictive power for your ML task.

Finally, Upgini will return a dataset enriched with only the relevant components of LLM embeddings.

If you already have features or other external data sources, you can specifically search for new datasets and features that only provide accuracy gains "on top" of them.

Just leave all these existing features in the labeled training dataset and the Upgini library automatically uses them during the feature search process and as a baseline ML model to calculate accuracy metric uplift. Only features that improve accuracy will be returned.

You can validate the robustness of external features on an out-of-time dataset using the eval_set parameter:

# load train dataset
train_df = pd.read_csv("train.csv")
train_ids_and_features = train_df.drop(columns="label")
train_label = train_df["label"]

# load out-of-time validation dataset
eval_df = pd.read_csv("validation.csv")
eval_ids_and_features = eval_df.drop(columns="label")
eval_label = eval_df["label"]
# create FeaturesEnricher
enricher = FeaturesEnricher(search_keys={"registration_date": SearchKey.DATE})

# now we fit WITH eval_set parameter to calculate accuracy metrics on Out-of-time dataset.
# the output will contain quality metrics for both the training data set and
# the eval set (validation OOT data set)
enricher.fit(
  train_ids_and_features,
  train_label,
  eval_set = [(eval_ids_and_features, eval_label)]
)

• Same data schema as for search initialization X dataset
• Pandas dataframe representation

The out-of-time dataset can be without labels. There are 3 options to pass out-of-time without labels:

enricher.fit(
  train_ids_and_features,
  train_label,
  eval_set = [
    (eval_ids_and_features_1,),  # A tuple with 1 element
    (eval_ids_and_features_2, None),  # None as labels
    (eval_ids_and_features_3, [np.nan] * len(eval_ids_and_features_3)),  # List or Series of the same size as eval X
  ]
)

FeaturesEnricher supports Population Stability Index (PSI) calculation on eval_set to evaluate feature stability over time. You can control this behavior using stability parameters in fit and fit_transform methods:

enricher = FeaturesEnricher(
    search_keys={"registration_date": SearchKey.DATE}
)

# Control feature stability during fit
enricher.fit(
    X, y, 
    stability_threshold=0.2,  # PSI threshold: features with PSI above this value will be dropped
    stability_agg_func="max"  # Aggregation function for stability values: "max", "min", "mean"
)

# Same parameters work for fit_transform
enriched_df = enricher.fit_transform(
    X, y,
    stability_threshold=0.1,   # Stricter threshold for more stable features
    stability_agg_func="mean"  # Use mean aggregation instead of max
)

Stability parameters:

• stability_threshold (float, default=0.2): PSI threshold value. Features with PSI above this threshold will be excluded from the final feature set. Lower values mean stricter stability requirements.
• stability_agg_func (str, default="max"): Function to aggregate PSI values across time intervals. Options: "max" (most conservative), "min" (least conservative), "mean" (balanced approach).

PSI (Population Stability Index) measures how much feature distribution changes over time. Lower PSI values indicate more stable features, which are generally more reliable for production ML models. PSI is calculated on the eval_set, which should contain the most recent dates relative to the training dataset.

FeaturesEnricher can be initialized with additional string parameter loss. Depending on the ML task, you can use the following loss functions:

• regression: regression, regression_l1, huber, poisson, quantile, mape, gamma, tweedie;
• binary: binary;
• multiclass: multiclass, multiclassova.

For instance, if your target variable has a Poisson distribution (count of events, number of customers in the shop and so on), you should try to use loss="poisson" to improve quality of feature selection and get better evaluation metrics.

Usage example:

enricher = FeaturesEnricher(
	search_keys={"date": SearchKey.DATE},
	loss="poisson",
    	model_task_type=ModelTaskType.REGRESSION
)
enriched_dataframe.fit(X, y)

fit, fit_transform, transform and calculate_metrics methods of FeaturesEnricher can be used with the exclude_features_sources parameter to exclude Trial or Paid features from Premium data sources:

enricher = FeaturesEnricher(
  search_keys={"subscription_activation_date": SearchKey.DATE}
)
enricher.fit(X, y, calculate_metrics=False)
trial_features = enricher.get_features_info()[enricher.get_features_info()["Feature type"] == "Trial"]["Feature name"].values.tolist()
paid_features = enricher.get_features_info()[enricher.get_features_info()["Feature type"] == "Paid"]["Feature name"].values.tolist()
enricher.calculate_metrics(exclude_features_sources=(trial_features + paid_features))
enricher.transform(X, exclude_features_sources=(trial_features + paid_features))

Upgini has autodetection of search keys enabled by default. To turn off use autodetect_search_keys=False:

enricher = FeaturesEnricher(
   search_keys={"date": SearchKey.DATE},
   autodetect_search_keys=False,
)

enricher.fit(X, y)

Upgini detects rows with target outliers for regression tasks. By default such rows are dropped during metrics calculation. To turn off the removal of target‑outlier rows, use the remove_outliers_calc_metrics=False parameter in the fit, fit_transform, or calculate_metrics methods:

enricher = FeaturesEnricher(
   search_keys={"date": SearchKey.DATE},
)

enricher.fit(X, y, remove_outliers_calc_metrics=False)

Upgini attempts to generate features for email, date and datetime search keys. By default this generation is enabled. To disable it use the generate_search_key_features parameter of the FeaturesEnricher constructor:

enricher = FeaturesEnricher(
  search_keys={"date": SearchKey.DATE},
  generate_search_key_features=False,
)

Register and get a free API key for exclusive data sources and features: 600M+ phone numbers, 350M+ emails, 2^32 IP addresses

You may publish ANY data which you consider as royalty‑ or license‑free (Open Data) and potentially valuable for ML applications for community usage:

1. Please Sign Up here
2. Copy Upgini API key from your profile and upload your data from the Upgini Python library with this key:

import pandas as pd
from upgini.metadata import SearchKey
from upgini.ads import upload_user_ads
import os
os.environ["UPGINI_API_KEY"] = "your_long_string_api_key_goes_here"
#you can define a custom search key that might not yet be supported; just use SearchKey.CUSTOM_KEY type
sample_df = pd.read_csv("path_to_data_sample_file")
upload_user_ads("test", sample_df, {
    "city": SearchKey.CUSTOM_KEY,
    "stats_date": SearchKey.DATE
})

3. After data verification, search results on community data will be available in the usual way.

Please note that we are still in beta. Requests and support, in preferred order

❗Please try to create bug reports that are:

• reproducible - include steps to reproduce the problem.
• specific - include as much detail as possible: which Python version, what environment, etc.
• unique - do not duplicate existing opened issues.
• scoped to a Single Bug - one bug per report.

We are not a large team, so we probably won't be able to:

• implement smooth integration with the most common low-code ML libraries and platforms (PyCaret, H2O AutoML, etc.)
• implement all possible data verification and normalization capabilities for different types of search keys And we need some help from the community!

So, we'll be happy about every pull request you open and every issue you report to make this library even better. Please note that it might sometimes take us a while to get back to you. For major changes, please open an issue first to discuss what you would like to change.

Some convenient ways to start contributing are:
⚙️ Open in Visual Studio Code You can remotely open this repo in VS Code without cloning or automatically clone and open it inside a docker container.
⚙️ Gitpod You can use Gitpod to launch a fully functional development environment right in your browser.

• Simple sales prediction template notebook
• Full list of Kaggle Guides & Examples
• Project on PyPI
• More perks for registered users

^{😔 Found typo or a bug in code snippet? Our bad! Please report it here}