🛠️ Feature Types Overview

Making Data ML-Ready

KDP makes feature processing intuitive and powerful by transforming your raw data into the optimal format for machine learning.

💪 Feature Types at a Glance

Feature Type	What It's For	Processing Magic
🔢 Numerical	Continuous values like age, income, scores	Normalization, scaling, embeddings, distribution analysis
🏷️ Categorical	Discrete values like occupation, product type	Embeddings, one-hot encoding, vocabulary management
📝 Text	Free-form text like reviews, descriptions	Tokenization, embeddings, sequence handling
📅 Date	Temporal data like signup dates, transactions	Component extraction, cyclical encoding, seasonality
➕ Cross Features	Feature interactions	Combined embeddings, interaction modeling
🔍 Passthrough	Pre-processed data, custom vectors	No modification, type casting only

🚀 Getting Started

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

from kdp import PreprocessingModel, FeatureType

# ✨ Quick and easy feature definition
features = {
    # 🔢 Numerical features - different processing strategies
    "age": FeatureType.FLOAT_NORMALIZED,        # 📊 [0,1] range normalization
    "income": FeatureType.FLOAT_RESCALED,       # 📈 Standard scaling
    "transaction_count": FeatureType.FLOAT,     # 🧮 Default normalization (same as FLOAT_NORMALIZED)

    # 🏷️ Categorical features - automatic encoding
    "occupation": FeatureType.STRING_CATEGORICAL,      # 👔 Job titles, roles
    "education_level": FeatureType.INTEGER_CATEGORICAL, # 🎓 Education codes

    # 📝 Text and dates - specialized processing
    "product_review": FeatureType.TEXT,         # 💬 Customer feedback
    "signup_date": FeatureType.DATE,            # 📆 User registration date

    # 🔍 Passthrough feature - use without any processing
    "embedding_vector": FeatureType.PASSTHROUGH # 🔄 Pre-processed data passes directly to output
}

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

⭐ Why Strong Feature Types Matter

🎯

Optimized Processing

Each feature type gets specialized handling for better ML performance

🐛

Reduced Errors

Catch type mismatches early in development, not during training

📝

Clearer Code

Self-documenting feature definitions make your code more maintainable

⚡

Enhanced Performance

Type-specific optimizations improve preprocessing speed

📚 Feature Type Documentation

🔢

Numerical Features

Handle continuous values with advanced normalization and distribution-aware processing

🏷️

Categorical Features

Process discrete categories with smart embedding techniques and vocabulary management

📝

Text Features

Work with free-form text using tokenization, embeddings, and sequence handling

📅

Date Features

Extract temporal patterns from dates with component extraction and cyclical encoding

➕

Cross Features

Model feature interactions with combined embeddings and interaction modeling

🔍

Passthrough Features

Include unmodified data or pre-computed features directly in your model

👨‍💻 Advanced Feature Configuration

For more control, use specialized feature classes:

from kdp.features import NumericalFeature, CategoricalFeature, TextFeature, DateFeature, PassthroughFeature
import tensorflow as tf

# 🔧 Advanced feature configuration
features = {
    # 💰 Numerical with advanced embedding
    "income": NumericalFeature(
        name="income",
        feature_type=FeatureType.FLOAT_RESCALED,
        use_embedding=True,
        embedding_dim=32
    ),

    # 🏪 Categorical with hashing
    "product_id": CategoricalFeature(
        name="product_id",
        feature_type=FeatureType.STRING_CATEGORICAL,
        max_tokens=10000,
        category_encoding="hashing"
    ),

    # 📋 Text with custom tokenization
    "description": TextFeature(
        name="description",
        max_tokens=5000,
        embedding_dim=64,
        sequence_length=128,
        ngrams=2
    ),

    # 🗓️ Date with cyclical encoding
    "purchase_date": DateFeature(
        name="purchase_date",
        add_day_of_week=True,
        add_month=True,
        cyclical_encoding=True
    ),

    # 🧠 Passthrough feature
    "embedding": PassthroughFeature(
        name="embedding",
        dtype=tf.float32
    )
}

💡 Pro Tips for Feature Definition

1

Start Simple

Begin with basic FeatureType definitions

2

Add Complexity Gradually

Refactor to specialized feature classes when needed

3

Combine Approaches

Mix distribution-aware, attention, embeddings for best results

4

Check Distributions

Review your data distribution before choosing feature types

5

Experiment with Types

Sometimes a different encoding provides better results

6

Consider Passthrough

Use passthrough features for pre-processed data or custom vectors

📊 Model Architecture Diagrams

KDP creates optimized preprocessing architectures based on your feature definitions. Here are examples of different model configurations:

🔄 Basic Feature Combinations

When combining numerical and categorical features:

🌟 All Feature Types Combined

KDP can handle all feature types in a single model:

🔋 Advanced Configurations

✨ Tabular Attention

Enhance feature interactions with tabular attention:

🔄 Transformer Blocks

Process categorical features with transformer blocks:

🧠 Feature MoE (Mixture of Experts)

Specialized feature processing with Mixture of Experts:

📤 Output Modes

KDP supports different output modes for your preprocessed features:

🔗 Concatenated Output

📦 Dictionary Output

---

← Architecture Numerical Features →