🔢 Numerical Features

Transform your continuous data like age, income, or prices into powerful feature representations

📋 Quick Overview

Numerical features are the backbone of most machine learning models. KDP provides multiple ways to handle them, from simple normalization to advanced neural embeddings.

🎯 Types and Use Cases

Feature Type	What it does	Output width	When to use
`FLOAT_NORMALIZED`	Standardised to zero mean and unit variance, using the mean and variance from the statistics pass.	1	The sensible default for most numeric columns.
`FLOAT`	Identical to `FLOAT_NORMALIZED`.	1	An alias; there is no separate behaviour.
`FLOAT_RESCALED`	Multiplied by `scale`, which defaults to 1.0 — that is, unchanged.	1	Only useful when you pass `scale` yourself. See the warning below.
`FLOAT_DISCRETIZED`	One-hot over `num_bins` equal-width bins derived from the statistics.	`num_bins` (default 10)	When groups of values carry meaning rather than the magnitude.

🚀 Basic Usage

The simplest way to define numerical features is with the FeatureType enum:

from kdp import PreprocessingModel, FeatureType

# ✨ Quick numerical feature definition
features = {
    "age": FeatureType.FLOAT_NORMALIZED,          # 🧓 zero mean, unit variance
    "income": FeatureType.FLOAT_NORMALIZED,       # 💰 also standardised (see the warning above)
    "transaction_count": FeatureType.FLOAT,       # 🔢 alias for FLOAT_NORMALIZED
    "rating": FeatureType.FLOAT_DISCRETIZED       # ⭐ one-hot over 10 bins
}

# 🏗️ Create your preprocessor
preprocessor = PreprocessingModel(
    path_data="customer_data.csv",
    features_specs=features
)

🧠 Advanced Configuration

For more control, use the NumericalFeature class:

from kdp.features import NumericalFeature

features = {
    # 🧓 Simple example with enhanced configuration
    "age": NumericalFeature(
        name="age",
        feature_type=FeatureType.FLOAT_NORMALIZED,
        use_embedding=True,                 # 🔄 Create neural embeddings
        embedding_dim=16,                   # 📏 Size of embedding
        preferred_distribution="normal"      # 📊 Hint about distribution
    ),

    # 💰 Financial data example
    "transaction_amount": NumericalFeature(
        name="transaction_amount",
        feature_type=FeatureType.FLOAT_RESCALED,
        use_embedding=True,
        embedding_dim=32,
        preferred_distribution="heavy_tailed"
    ),

    # ⏳ Custom binning example
    "years_experience": NumericalFeature(
        name="years_experience",
        feature_type=FeatureType.FLOAT_DISCRETIZED,
        num_bins=5                          # 📏 Number of bins
    )
}

⚙️ Key Configuration Options

Parameter	Description	Default	Suggested Range
`feature_type`	🏷️ Base feature type	`FLOAT_NORMALIZED`	Choose from 4 types
`use_embedding`	🧠 Enable neural embeddings	`False`	`True`/`False`
`embedding_dim`	📏 Dimensionality of embedding	8	4-64
`preferred_distribution`	📊 Hint about data distribution	`None`	"normal", "log_normal", etc.
`num_bins`	🔢 Bins for discretization	10	5-100

🔥 Power Features

📊

Distribution-Aware Processing

Let KDP automatically detect and handle distributions:

# ✨ Enable distribution-aware processing for all numerical features
preprocessor = PreprocessingModel(
    features_specs=features,
    use_distribution_aware=True      # 🔍 Enable distribution detection
)

  </div>
</div>

🧠

Advanced Numerical Embeddings

Using advanced numerical embeddings:

from kdp import FeatureType, NumericalFeature, PreprocessingModel

# Configure numerical embeddings
preprocessor = PreprocessingModel(
    features_specs={
        "income": NumericalFeature(
            name="income",
            feature_type=FeatureType.FLOAT_RESCALED,
            use_embedding=True,
            embedding_dim=32,
            preferred_distribution="log_normal"
        )
    }
)

  </div>
</div>

💼 Real-World Examples

💰

Financial Analysis

from kdp import FeatureType, NumericalFeature, PreprocessingModel

# 📈 Financial metrics with appropriate processing
preprocessor = PreprocessingModel(
    features_specs={
        "income": NumericalFeature(
            name="income",
            feature_type=FeatureType.FLOAT_RESCALED,
            preferred_distribution="log_normal"   # 📉 Log-normal distribution
        ),
        "credit_score": NumericalFeature(
            name="credit_score",
            feature_type=FeatureType.FLOAT_NORMALIZED,
            use_embedding=True,
            embedding_dim=16
        ),
        "debt_ratio": NumericalFeature(
            name="debt_ratio",
            feature_type=FeatureType.FLOAT_NORMALIZED,
            preferred_distribution="bounded"      # 📊 Bounded between 0 and 1
        )
    },
    use_distribution_aware=True                   # 🧠 Smart distribution handling
)

</div>

🔌

Sensor Data

from kdp import FeatureType, NumericalFeature, PreprocessingModel

# 📡 Processing sensor readings
preprocessor = PreprocessingModel(
    features_specs={
        "temperature": NumericalFeature(
            name="temperature",
            feature_type=FeatureType.FLOAT_RESCALED,
            use_embedding=True,
            embedding_dim=16
        ),
        "humidity": NumericalFeature(
            name="humidity",
            feature_type=FeatureType.FLOAT_NORMALIZED,
            preferred_distribution="bounded"      # 💧 Bounded between 0 and 100
        ),
        "pressure": NumericalFeature(
            name="pressure",
            feature_type=FeatureType.FLOAT_RESCALED,
            use_embedding=True,
            embedding_dim=16
        )
    }
)

</div>

💡 Pro Tips

📊

Understand Your Data Distribution

Use FLOAT_NORMALIZED when your data has clear bounds (e.g., 0-100%)
Use FLOAT_RESCALED when your data has outliers (e.g., income, prices)
Use FLOAT_DISCRETIZED when your values naturally form groups (e.g., age groups)

🧠

Consider Neural Embeddings for Complex Relationships

Enable when a simple scaling doesn't capture the pattern
Increase embedding dimensions for more complex patterns (16→32→64)

🔍

Let KDP Handle Distribution Detection

Enable use_distribution_aware=True and let KDP automatically choose
This is especially important for skewed or multi-modal distributions

📏

Custom Bin Boundaries

Use num_bins parameter to control discretization granularity
More bins = finer granularity but more parameters to learn

🔗 Related Topics

📊

Distribution-Aware Encoding

Smart numerical handling

🧠

Advanced Numerical Embeddings

Neural representations

🎯

Feature Selection

Finding important features

🧮 Types of Numerical Features

KDP supports different types of numerical features, each with specialized processing:

🔄

FLOAT

Basic floating-point features with default normalization

<div class="type-card">
  <span class="type-icon">📏</span>
  <h3>FLOAT_NORMALIZED</h3>
  <p>Values normalized to the [0,1] range using min-max scaling</p>
</div>

<div class="type-card">
  <span class="type-icon">⚖️</span>
  <h3>FLOAT_RESCALED</h3>
  <p>Values rescaled using standardization (mean=0, std=1)</p>
</div>

<div class="type-card">
  <span class="type-icon">📊</span>
  <h3>FLOAT_DISCRETIZED</h3>
  <p>Continuous values binned into discrete buckets</p>
</div>

📊 Architecture Diagrams

📏 Normalized Numerical Feature

Below is a visualization of a model with a normalized numerical feature:

⚖️ Rescaled Numerical Feature

Below is a visualization of a model with a rescaled numerical feature:

📊 Discretized Numerical Feature

Below is a visualization of a model with a discretized numerical feature:

🧠 Advanced Numerical Embeddings

When using advanced numerical embeddings, the model architecture looks like this:

← Feature Overview Categorical Features →

FilesExpand file tree

numerical-features.md

Latest commit

History

numerical-features.md

File metadata and controls

🔢 Numerical Features

Transform your continuous data like age, income, or prices into powerful feature representations

📋 Quick Overview

🎯 Types and Use Cases

🚀 Basic Usage

🧠 Advanced Configuration

⚙️ Key Configuration Options

🔥 Power Features

Distribution-Aware Processing

Advanced Numerical Embeddings

💼 Real-World Examples

Financial Analysis

Sensor Data

💡 Pro Tips

Understand Your Data Distribution

Consider Neural Embeddings for Complex Relationships

Let KDP Handle Distribution Detection

Custom Bin Boundaries

🔗 Related Topics

Distribution-Aware Encoding

Advanced Numerical Embeddings

Feature Selection

🧮 Types of Numerical Features

FLOAT

📊 Architecture Diagrams

📏 Normalized Numerical Feature

⚖️ Rescaled Numerical Feature

📊 Discretized Numerical Feature

🧠 Advanced Numerical Embeddings