SAS^® Factory Miner Features

Data access & preparation

• Access data sources registered in SAS Metadata Server, including SAS data sets, database tables and Hadoop files.
• Interactively assign data source metadata – such as variable roles, levels or order – or use automated settings to share variable settings across projects.
  
• Define segments in your data for stratified modeling.
• Assess data issues with automated data profiling and interactive variable distribution graphs.
  
• Filter your analytical input data:
  
  • Winsorized.
  • Standard deviations.
  • Trimmed.
  • Rare values.
• Apply transformations to your data for better models:
  • Log.
  • Log 10.
  • Square root.
    
  • Inverse.
    
  • Square.
    
  • Exponential.
  • Centering.
    
  • Standardized.
    
  • Range.
    
  • Bucket.
  • Pseudo-quantile.
  • Optimal binning.
    
  • Principal component analysis.
    
• Clean your data with statistical and machine learning imputation methods:
  • Mean.
  • Min.
  • Max.
  • Median.
  • Midrange.
  • Constant.
  • Count-based.
  • Distribution-based.
    

Customizable supervised learning templates

• Interactively build custom templates that include models and the following processing steps:
  
  • Filtering.
  • Principal components.
  • Imputation.
  • Transformations.
  • Supervised and unsupervised variable selection.
• Create your own model templates.
• Edit any data preparation or model parameters and save as customized template.
  
• Share model templates across projects and users.

Self-service machine learning techniques

• Build models using the following techniques:
  • Bayesian networks.
  • Decision trees.
  • Gradient boosting.
  • Neural networks.
  • Random forests.
  • Support vector machines.
  • Generalized linear models.
  • Linear regression.
  • Logistic regression.
• Interactively view model-specific results.

Champion model identification

• Champion models are automatically selected for each segment using selectable criteria:
  • Kolmogorov-Smirnov.
  • Lift and cumulative lift.
  • Gain and cumulative gain.
  • Misclassification rate.
  • Percent captured event.
  • Average percent captured event.
  • Average square error.
• Override system-selected models and manually identify your champion model.
• Interactively compare and assess models within a segment and across multiple segments.

Model exceptions identification

• View reports that highlight model performance exceptions.
• Easily identify and drill into details for underperforming models.
• Modify default settings for each model template.

Model tracking & reporting

• Generate summary reports that contain model results, significant variables and model settings.
  
• Share reports via PDFs and RTFs.
  

Model retraining

• Retrain existing model templates on new data sets.
• Track model-build assessment statistics across retraining iterations.
• Longitudinal model performance degradation reports.
• Automatically retrain models in batch using REST endpoints.

Flexible model management & deployment

• Automatically generate SAS score code for all model templates.
• Register models to SAS Model Manager for centralized model deployment and management (requires SAS Model Manager).
• Deploy models in database and in Hadoop using SAS Scoring Accelerator (requires SAS Scoring Accelerator).
  

Scalable processing

• Train models using multithreaded procedures on SAS servers to take advantage of multicore servers.
  
• Train models using asynchronous processes via SAS Grid Manager for workload balancing and scheduling (requires SAS Grid Manager).
• Train models in memory using SAS High-Performance Data Mining on database appliances (Oracle, Teradata, Greenplum, SAP HANA) or on Hadoop (requires SAS High-Performance Data Mining).