Skip to content

Models Supported in SmartML

SmartML supports a fixed set of classification and regression models covering:

  • Classical machine learning baselines
  • Tree-based ensemble methods
  • Deep learning models for tabular data
  • Specialized neighborhood-based models

All models are instantiated with fixed defaults to ensure fair and reproducible benchmarking.

Model availability depends on installed optional dependencies.
Unavailable models are automatically excluded at runtime.

Classification Models

Baseline Models

These models provide strong and interpretable baselines for tabular classification.

  • Logistic Regression
  • Linear classifier with probabilistic outputs
  • Default solver and regularization from scikit-learn
  • No hyperparameter tuning

  • Used as the minimal baseline

  • Support Vector Classifier (SVC)

  • Margin-based classifier
  • Default kernel and parameters from scikit-learn
  • Feature scaling applied internally

  • No kernel or C-value search

  • K-Nearest Neighbors (KNN)

  • Distance-based classifier
  • Default number of neighbors
  • Feature scaling applied

  • Sensitive to feature space geometry

  • Naive Bayes

  • Probabilistic classifier
  • Uses standard distribution assumptions
  • No smoothing or prior tuning

Tree-Based Ensemble Models

Tree ensembles are widely used for tabular data due to robustness and performance.

  • Random Forest
  • 100 trees
  • Parallel execution enabled
  • Fixed random seed (42)

  • No depth or feature tuning

  • Extra Trees

  • 100 trees
  • High randomization in split selection
  • Parallel execution enabled

  • Fixed random seed (42)

  • LightGBM

  • 100 boosting iterations
  • CPU execution
  • Parallel training enabled
  • Logging disabled

  • Fixed random seed (42)

  • XGBoost

  • 100 boosting iterations
  • Histogram-based tree method
  • CPU execution
  • Parallel training enabled
  • Logging disabled

  • Fixed random seed (42)

  • CatBoost

  • 100 boosting iterations
  • Native categorical handling
  • Silent training mode
  • Fixed random seed (42)

Deep Learning Models (Tabular)

Deep models are included for comparison, not tuning or production usage.
All deep models use conservative, deterministic defaults.

Common characteristics:

  • CPU execution
  • Fixed number of training epochs
  • Deterministic training enabled
  • No logging or progress bars

Models included:

  • MLP
  • Fully connected neural network
  • Two hidden layers (256 → 128)
  • Dropout enabled

  • Fixed epoch count

  • FT-Transformer

  • Transformer-based architecture for tabular data
  • Fixed number of attention heads
  • Fixed hidden dimension

  • No architecture search

  • TabTransformer

  • Transformer with categorical embeddings
  • Default embedding and attention configuration

  • Deterministic execution

  • SAINT

  • Self-attention based tabular model
  • Default attention configuration

  • Fixed training setup

  • TabNet

  • Sparse attention with feature selection
  • Fixed number of decision steps

  • Deterministic seed

  • NODE

  • Neural Oblivious Decision Ensembles
  • Library defaults

  • No depth or ensemble tuning

  • NAM

  • Neural Additive Model
  • Explainability-focused architecture

  • Default configuration only

  • GrowNet

  • Boosting-inspired neural architecture
  • Library defaults

  • CPU execution

  • Modern NCA

  • Neural neighborhood-based classifier
  • Default neighborhood configuration
  • Deterministic execution

Specialized Models

  • SmartKNN (Classification)
  • Optimized neighborhood-based classifier
  • Custom distance and candidate selection logic
  • Designed for fast inference and scalability
  • Used as a strong non-parametric baseline

Regression Models

Baseline Models

All baseline regression models apply standard feature scaling internally.

  • Linear Regression
  • Ordinary least squares
  • No regularization

  • Scaled inputs

  • Ridge Regression

  • L2 regularization
  • Default regularization strength

  • Scaled inputs

  • Lasso Regression

  • L1 regularization
  • Default sparsity strength

  • Scaled inputs

  • ElasticNet

  • Combination of L1 and L2 regularization
  • Default mixing ratio

  • Scaled inputs

  • Support Vector Regressor (SVR)

  • Default kernel and parameters
  • Scaled inputs

  • No kernel tuning

  • K-Nearest Neighbors Regressor

  • Fixed number of neighbors
  • Distance-based prediction
  • Scaled inputs

Tree-Based Ensemble Models

  • Random Forest Regressor
  • 100 trees
  • Parallel execution enabled

  • Fixed random seed (42)

  • Extra Trees Regressor

  • 100 trees
  • Highly randomized splits
  • Parallel execution enabled

  • Fixed random seed (42)

  • LightGBM Regressor

  • 100 boosting iterations
  • CPU execution
  • Logging disabled

  • Fixed random seed (42)

  • XGBoost Regressor

  • 100 boosting iterations
  • Histogram-based training
  • CPU execution

  • Fixed random seed (42)

  • CatBoost Regressor

  • 100 boosting iterations
  • Native categorical support
  • Silent mode
  • Fixed random seed (42)

Deep Learning Models (Tabular)

Deep regression models follow the same execution constraints as classification:

  • CPU-only execution
  • Deterministic behavior
  • Fixed training configuration
  • No hyperparameter search

Models included:

  • MLP Regressor
  • FT-Transformer Regressor
  • TabTransformer Regressor
  • SAINT Regressor
  • TabNet Regressor
  • NODE Regressor
  • NAM Regressor
  • DeepGBM Regressor
  • GrowNet Regressor

All use library defaults with fixed training limits.

Specialized Models

  • SmartKNN (Regression)
  • Optimized neighborhood-based regressor
  • Fast inference-oriented design
  • Deterministic execution

Model Registry Behavior

SmartML maintains separate registries for classification and regression.

  • Models are selected based on task type
  • Missing dependencies automatically disable affected models
  • Only available models are listed and executed
  • Model keys are normalized internally

Design Rationale

Model inclusion follows these principles:

  • Coverage of common tabular modeling approaches
  • Strong baselines over tuned variants
  • Fixed defaults for fairness
  • Deterministic and reproducible behavior

Hyperparameter tuning is intentionally excluded.