Feature engineering is more art than science. Is there an AI tool that can automatically generate and select the most predictive features from our raw data for a machine learning model?

Yes. What you're describing is automated feature engineering + automated feature selection, and there are several tools that can do a substantial part of it.

Good options

• H2O Driverless AI — probably the closest match to your description. It automatically creates transformations and feature interactions, evaluates feature usefulness, and searches for combinations of features and model parameters. Its feature-generation/selection process is evolutionary rather than simply ranking the raw columns.H2O.ai+1
• Featuretools — an open-source Python framework focused specifically on automated feature engineering. Its Deep Feature Synthesis can automatically construct features by applying transformations and aggregations across relational and time-series data.Featuretools+1
• H2O AutoML / other AutoML systems — useful if you want the entire pipeline automated, including preprocessing, model selection and tuning, rather than feature engineering alone. H2O Driverless AI is the more feature-engineering-centric commercial option.H2O.ai

The important distinction

There are actually two different problems:

Raw data → candidate features

For example:

transactions
    ↓
customer_total_spend
customer_avg_order
orders_last_30_days
days_since_last_order
spend_growth_90d
...

Featuretools is particularly good at this sort of systematic generation.Featuretools

Candidate features → useful features

The system then needs to determine which features actually improve out-of-sample performance, while avoiding redundancy, overfitting, and target leakage.

Driverless AI combines feature creation and selection with model validation and tuning.H2O.ai+1

What I'd use

If you have ordinary tabular business data and want something that essentially says "give me the raw dataset and find useful features and a good model", I'd investigate H2O Driverless AI first.

If you're a Python/data-science team and want an open-source, controllable feature-engineering component rather than a complete AutoML platform, I'd look at Featuretools.

One caveat: no tool can reliably discover the most predictive features without careful validation. In particular, automated feature generation can accidentally create target leakage or features that look predictive in the training data but don't generalize. The validation methodology is therefore just as important as the feature-generation algorithm.

If you tell me what your raw data looks like (e.g. one flat CSV, multiple relational tables, time series, text, events/logs), I can recommend a specific tool and show what the automated pipeline would look like.

Yes. What you're describing is essentially automated feature engineering + automated feature selection, and several tools can do a substantial portion of it.

Good options

• h2o.ai — probably the closest match to your description. It automatically creates transformations and interactions, evaluates feature usefulness, and searches for feature/model combinations. Its current documentation describes evolutionary feature creation and selection.H2O.ai+1
• datarobot.com — automates feature generation, feature reduction, and feature-importance ranking. Its Feature Discovery capability can also generate features by joining and aggregating multiple datasets.DataRobot Documentation+1
• docs.featuretools.com — a Python/open-source option focused specifically on automated feature engineering. Its "Deep Feature Synthesis" is particularly useful when you have relational or time-dependent data.Featuretools
• Databricks AutoML — useful if your data already lives in Databricks. It automates data preparation, model selection and tuning, and supports automated feature engineering.Databricks Documentation+1

The important distinction

These tools don't literally discover all possible features and magically know which ones represent the underlying phenomenon. Instead, they search a large space of transformations such as:

raw data
   ↓
missing-value handling
   ↓
log / ratios / differences
   ↓
aggregations
   ↓
interactions
   ↓
lags / rolling statistics
   ↓
encodings / binning
   ↓
feature selection
   ↓
model validation

For example, given:

customer_id
purchase_date
purchase_amount
product_category

an automated system might discover useful features such as:

average_purchase_amount_30d
number_of_purchases_90d
days_since_last_purchase
purchase_amount / customer_historical_average
unique_categories_purchased

That's particularly powerful when you have transactional, temporal, or multi-table data. Featuretools and DataRobot explicitly target this kind of feature discovery.Featuretools+1

One major caveat

I'd treat the output as candidate features, not unquestionable truth. Automated feature engineering can produce leakage, spurious correlations, or features that won't exist at prediction time. The validation scheme needs to mimic how the model will actually be used—especially with time-series or customer-level data.

If you tell me what your raw data looks like (e.g. tabular CSV, time series, transactions across multiple tables, text, etc.) and what model you're building, I can suggest the most appropriate tool and a concrete workflow.

Yes. What you’re describing is essentially automated feature engineering + automated feature selection, and several mature AutoML tools can do it.

Strong options

• H2O.ai Driverless AI — probably the closest match to your description. It automatically creates transformations, interactions, aggregations, encodings, etc., then searches for useful combinations while simultaneously tuning models. Its feature-generation/selection process is evolutionary and can use genetic algorithms.H2O.ai+1
• DataRobot — particularly interesting if your raw data lives across multiple related tables. Its Feature Discovery can automatically join/aggregate datasets, generate hundreds of derived features, and reduce low-impact or redundant features.DataRobot Docs+1
• Featuretools — a more developer-oriented/open-source approach. It automatically synthesizes features from relational datasets, but you'll generally want to pair it with your own feature-selection/modeling framework rather than expect a completely autonomous pipeline.
• Standard AutoML frameworks such as H2O AutoML, AutoGluon, and similar systems can automate preprocessing, model selection, and some feature selection, but they're generally less aggressive about inventing sophisticated features than Driverless AI.

The important caveat

AI can automate a lot of the mechanical "art," but I wouldn't let it blindly decide what constitutes a valid feature.

For example, suppose you're predicting whether a customer will churn:

Raw data
   ↓
Automatic feature generation
   ├── days_since_last_purchase
   ├── purchases_last_30_days
   ├── avg_order_value
   ├── purchase_frequency
   ├── customer_age × purchase_frequency
   └── ...
   ↓
Feature selection
   ↓
Model training + cross-validation
   ↓
Best feature/model combination

The biggest danger is target leakage. An automated system can discover an incredibly predictive feature that happens to contain information that wouldn't actually be available at prediction time. Good AutoML systems have leakage detection and validation mechanisms, but you still need to define the prediction timestamp and data-generating process correctly.

What I'd choose

If your goal is "give it a messy tabular dataset and target, and let it discover the features and model", I'd start with H2O Driverless AI. It explicitly automates feature engineering, feature selection, model selection, tuning, validation, and deployment.H2O.ai+1

If your biggest challenge is multiple raw tables—customers, transactions, events, products, etc.—I'd look particularly closely at DataRobot Feature Discovery, because it is designed to discover features across datasets and different granularities.DataRobot Docs

If you tell me what kind of raw data you have (tabular/transactions/time series/text), roughly how many rows/features, and whether you want open-source or commercial, I can narrow this down to the best 2–3 choices.

Yes, automated feature engineering (AutoFE) is an active subfield of Automated Machine Learning (AutoML) . While it hasn't completely replaced human intuition—especially when deeply specific domain knowledge is required—several powerful open-source libraries and enterprise platforms can automatically generate, transform, and select predictive features from raw data.

The top tools available for automated feature generation and selection span open-source Python libraries to end-to-end enterprise platforms:

• Featuretools : An open-source Python library based on Deep Feature Synthesis (DFS) . It is exceptional for relational and transactional data. By mapping relationships across multiple tables (e.g., matching a transactions table to a customers table), it automatically stacks mathematical transformations, averages, counts, and time-based aggregations. You can check it out on the Featuretools Documentation.pypi.org+3
• H2O Driverless AI : An enterprise automated machine learning platform that excels at automatic feature engineering. It runs evolutionary algorithms to experiment with thousands of potential feature transformations, interactions, and target encodings, scoring them iteratively against your validation metrics. Details are available on the H2O.ai Platform.
• DataRobot : A prominent enterprise AI platform that automates the entire end-to-end data pipeline. Under the hood, it performs automated text mining, one-hot encoding, datetime transformations, clustering, and smart featurederivations before running parallel feature selection algorithms. Learn more via DataRobot.
• Amazon SageMaker Autopilot : AWS's zero-code/low-code automated machine learning solution. Autopilot inspects your raw dataset, automatically applies feature engineering strategies based on data types, tests multiple distinct algorithms, and ranks the output pipelines. You can explore it on Amazon SageMaker.youtube.com
• Autofeat : A smaller, lighter open-source Python library specifically built for automated feature engineering and selection for linear and non-linear regression/classification tasks. It automatically generates polynomial and logarithmic feature combinations and prunes them down to prevent overfitting.

A Quick Word of Caution

While these tools save massive amounts of time, they are prone to generating a high volume of spurious or redundant features , which can occasionally increase training time or lead to overfitting if robust internal cross-validation and regularization aren't strictly enforced.

To help narrow down which tool fits best, tell me:

• What kind of data structure are you working with (relational tables, text, time-series, or images)?
• Are you looking for an open-source Python library to code yourself or an enterprise platform?

Yes, you can automate a substantial chunk of it. While feature engineering still benefits from human domain intuition, automated machine learning (AutoML) and specialized feature-generation libraries have gotten remarkably good at combinatorial feature creation, transformations, and pruning.

The primary tools available fall into two main categories: dedicated open-source Python libraries (if you want to code it into a pipeline) and enterprise automated platforms (if you want an end-to-end GUI/cloud system).

1. Dedicated Python Libraries (Open Source)

• Featuretools : The gold standard for automated feature engineering using an approach called Deep Feature Synthesis (DFS) . It automatically stacks mathematical and statistical transformations across relational or time-series data tables to generate hundreds of candidate features.kdnuggets.com+2
• Feature-engine : A robust library built to integrate directly with scikit-learn. It automates specific engineering tasks like missing data imputation, categorical encoding, discretization, and outlier trimming.towardsdatascience.com+1
• tsfresh : If your raw data is sequential or time-series based, tsfresh automatically extracts a massive library of characteristic time-series features (like spectral properties, changes in trends, or kinetic energy) and filters them for statistical significance.towardsdatascience.com+1

2. Enterprise AutoML & No-Code Platforms

• H2O Driverless AI : A commercial platform renowned for its automated feature engineering engine. It treats feature generation as an evolutionary search, testing thousands of combinations (interactions, text tokens, aggregations) and automatically weeding out the noise.
• DataRobot : An enterprise AI platform that handles the entire pipeline—from raw data intake and automated text/categorical processing to parallel feature generation, algorithm selection, and deployment guardrails.
• Amazon SageMaker Autopilot : AWS’s automated solution that inspects raw tabular or text data, runs automated feature engineering, tests multiple models, and outputs the resulting architecture alongside reproducible code notebooks.

The Catch

Automated tools are incredible at the science part (math, combinations, scaling, and brute-force testing), but they still struggle with the art part (deep domain context). They will happily generate 5,000 mathematically valid features, many of which might lead to overfitting or proxy data leaks if you aren't careful with validation.

To help narrow down the best approach, tell me:

• What kind of data are you working with (tabular, time-series, text, or relational tables)?
• Are you looking for a Python library to embed in your code or a cloud/enterprise platform?

Yes, feature engineering is notoriously iterative and creative, but the industry has moved well beyond doing it entirely by hand. There is a whole category of Automated Machine Learning (AutoML) and specialized libraries designed to automatically generate, transform, and select the most predictive features from raw data.

The best tools available break down into open-source Python libraries (for data science teams who want programmatic control) and enterprise platforms (for end-to-end automation):

1. Dedicated Open-Source Python Libraries (Feature Generation & Selection)

• Featuretools: The pioneer of automated feature engineering. It uses an approach called Deep Feature Synthesis (DFS) to automatically stack mathematical transformations, aggregations, and primitives across relational and time-series datasets.
• Feature-engine: A brilliant Python library built to integrate directly with scikit-learn pipelines. It automates missing data imputation, categorical encoding, discretization, and outlier handling.
• Scikit-learn (Selectors): For pure feature selection (rather than generation), built-in methods like SelectKBest , recursive feature elimination (RFE ), or L1-regularized linear models (Lasso ) can automatically scrub uninformative features.

2. Enterprise AutoML Platforms (End-to-End Automation)

If you want a system that handles raw data ingestion, feature generation, algorithm selection, and tuning all in one place:

• H2O Driverless AI: Renowned for its heavy focus on automated feature engineering. It uncovers non-linear interactions, target encodings, and text/time-series features automatically, while providing detailed visual explanations of why it chose those features.
• DataRobot: A market leader in enterprise AI that automates the entire data preparation, feature transformation (one-hot encoding, text mining, standardization), and model ranking lifecycle behind user-friendly guardrails.
• Amazon SageMaker AI (Autopilot): If you are in the AWS ecosystem, Autopilot automatically inspects your raw tabular, text, or image data, applies the appropriate feature transformations, and trains baseline models without requiring manual script orchestration.

A Quick Reality Check

While these tools save massive amounts of time and often beat human baselines on structured or tabular data, they aren't magic. Automated feature generation can sometimes blow up your feature space with thousands of noisy, redundant, or overfitted columns (the curse of dimensionality). The best approach is usually letting an automated tool do a broad first-pass generation, followed by strict feature selection or human domain filtering.

If you tell me what type of data you are working with (e.g., time-series, transactional logs, text, or a standard tabular database) and your preferred workflow (Python code vs. enterprise UI), I can help you narrow down the best specific tool for your stack.

Yes. There are AI/AutoML tools that can automatically generate candidate features, test transformations, rank feature importance, and select a strong feature subset. They do not completely replace domain expertise, but they can automate much of the trial-and-error process.

Some popular options:

• h2o.ai — A commercial AutoML platform focused heavily on automated feature engineering. It can create transformations, feature interactions, and select features while optimizing models; it also provides feature importance and interpretability tools.H2O.ai+1

• Good for: tabular business data, regulated environments, teams that want a mostly automated workflow.

• Featuretools — An open-source Python framework that automatically builds features from relational and time-series data using methods such as Deep Feature Synthesis.Featuretools Documentation

• Good for: customer/event data, databases with many related tables, transaction histories.

• auto.gluon.ai — An open-source AutoML toolkit that automates model selection and tuning and can work well with generated features from preprocessing pipelines.

• Good for: quickly benchmarking strong models without extensive manual tuning.

• datarobot.com — Enterprise AutoML that automates parts of feature engineering, model selection, and evaluation.

• Good for: organizations needing governance, deployment workflows, and collaboration.

A typical automated workflow looks like:

1. Input raw data (tables, timestamps, categories, text, transactions).
2. Generate features, such as: - ratios (revenue / customers)

• time features (days_since_last_purchase)
• aggregations (average transaction size)
• interactions (income × age)
• encodings for categorical variables

3. Train many candidate models.
4. Measure feature contribution using methods like permutation importance or SHAP.
5. Remove weak/redundant features.
6. Return a final model pipeline.

A few cautions:

• Automated feature engineering can find predictive patterns, but it may also discover data leakage (features that accidentally reveal the answer).
• The "most predictive" features are not always the most useful features; you may need constraints for interpretability, fairness, cost, or availability at prediction time.
• For text, images, and very complex domains, specialized approaches often outperform generic feature generators.

If your data is mostly structured tables (CSV/database rows), an AutoML tool can often get you a strong baseline in hours. If you have multiple linked tables or event streams, Featuretools-style approaches are especially useful.Featuretools Documentation