SPM | Salford Predictive Modeler

SALFORD PREDICTIVE MODELER

Features List

Salford Predictive Modeler® 8 General Features

Modeling Engine: CART® decision trees
Modeling Engine: TreeNet® gradient boosting
Modeling Engine: Random Forests® tree ensemble
Modeling Engine: MARS® nonlinear regression splines
Modeling Engine: GPS regularized regression (LASSO, Elastic Net, Ridge, etc.)
Modeling Engine: RuleLearner, incorporating TreeNet’s accuracy plus the interpretability of regression
Modeling Engine: ISLE model compression
70+ pre-packaged automation routines for enhanced model building and experimentation
Tools to relieve gruntwork, allowing the analyst to focus on the creative aspects of model development.
Open Minitab Worksheet (.MTW) functionality

CART® Features

Hotspot detection to discover the most important parts of the tree and the corresponding tree rules
Variable importance measures to understand the most important variables in the tree
Deploy the model and generate predictions in real-time or otherwise
User defined splits at any point in the tree
Differential lift (also called “uplift” or “incremental response”) modeling for assessing the efficacy of a treatment
Automation tools for model tuning and other experiments including
- Automatic recursive feature elimination for advanced variable selection
- Experiment with the prior probabilities to obtain a model that achieves better accuracy rates for the more important class
- Perform repeated cross validation
- Build CART models on bootstrap samples
- Build two linked models, where the first one predicts a binary event while the second one predicts a numeric value
- Discover the impact of different learning and testing partitions

MARS® Features

Graphically understand how variables affect the model response
Determine the importance of a variable or set of interacting variables
Deploy the model and generate predictions in real-time or otherwise
Automation tools for model tuning and other experiments including
- Automatic recursive feature elimination for advanced variable selection
- Automatically assess the impact of allowing interactions in the model
- Easily find the best minimum span value
- Perform repeated cross validation
- Discover the impact of different learning and testing partitions

TreeNet® Features

Graphically understand how variables affect the model response with partial dependency plots
Regression loss functions: least squares, least absolute deviation, quantile, Huber-M, Cox survival, Gamma, Negative Binomial, Poisson, and Tweedie
Classification loss functions: binary or multinomial
Differential lift (also called “uplift” or “incremental response”) modeling
Column subsampling to improve model performance and speed up the runtime.
Regularized Gradient Boosting (RGBOOST) to increase accuracy.
RuleLearner: build interpretable regression models by combining TreeNet gradient boosting and regularized regression (LASSO, Elastic Net, Ridge etc.)
ISLE: Build smaller, more efficient gradient boosting models using regularized regression (LASSO, Elastic Net, Ridge, etc.)
Variable Interaction Discovery Control
- Determine definitively whether or not interactions of any degree need to be included
- Control the interactions allowed or disallowed in the model with Minitab’s patented interaction control language
Discover the most important interactions in the model
Calibration tools for rare-event modeling
Automation tools for model tuning and other experiments including
- Automatic recursive feature elimination for advanced variable selection
- Experiment with different learn rates automatically
- Control the extent of interactions occurring in the model
- Build two linked models, where the first one predictions a binary event while the second one predicts a numeric value
- Find the best parameters in your regularized gradient boosting model
- Perform a stochastic search for the core gradient boosting parameters
- Discover the impact of different learning and testing partitions

Random Forests® Features

Use for classification, regression, or clustering
Outlier detection
Proximity heat map and multi-dimensional scaling for graphically determining clusters in classification problems (binary or multinomial)
Parallel Coordinates Plot for a better understanding of what levels of predictor values lead to a particular class assignment
Unsupervised learning: Random Forest creates the proximity matrix and hierarchical clustering techniques are then applied
Variable importance measures to understand the most important variables in the model
Deploy the model and generate predictions in real-time or otherwise
Automation tools for model tuning and other experiments including
- Automatic recursive feature elimination for advanced variable selection
- Easily fine tune the random subset size taken at each split in each tree
- Assess the impact of different bootstrap sample sizes
- Discover the impact of different learning and testing partitions