Developer Guide for Intel® Data Analytics Acceleration Library 2019 Update 5

Training and Prediction

Training and prediction algorithms in Intel® Data Analytics Acceleration Library (Intel® DAAL) include a range of popular machine learning algorithms. Unlike analysis algorithms, which are intended to characterize the structure of data sets, machine learning algorithms model the data. Modeling operates in two major stages:

  1. Training.

    At this stage, the algorithm estimates model parameters based on a training data set.

  2. Prediction or decision making.

    At this stage, the algorithm uses the trained model to predict the outcome based on new data.

The following major categories of training and prediction methods are available in Intel DAAL:

Training is typically a lot more computationally complex problem than prediction. Therefore, certain end-to-end analytics usage scenarios require that training and prediction phases are done on distinct devices, the training is done on more powerful devices, while prediction is done on smaller devices. Because smaller devices may have stricter memory footprint requirements, Intel DAAL separates Training, Prediction, and respective Model in three different class hierarchies to minimize the footprint.

An alternative to training your model with algorithms implemented in Intel DAAL is to build a trained model from pre-calculated model parameters, for example, coefficients β for Linear Regression. This enables you to use Intel DAAL only to get predictions based on the model parameters computed elsewhere.

The Model Builder class provides an interface for adding all the necessary parameters and building a trained model ready for the prediction stage.

The following schema illustrates the use of Model Builder class:
Model Builder class usage schema

The Model Builder class is implemented for Linear and Logistic Regression, SVM Classifier, Multi-Class Classifier, Gradient Boosted Trees Regression, Gradient Boosted Trees Classification, and for Decision Forest Classification.