AutoAI overview (Watson Machine Learning)

Choosing a compute configuration

When you configure your experiment, you can choose a compute configuration that supplies the computing resources for your experiment. The configuration that you choose governs data limits.

Notes:

• You can change how your training data is sampled from the Experiment Settings.
• If you are using a larger data source (up to 100 GB), the AutoAI experiment can train with all of the data by using ensembling and incremental learning.
• If you are connecting to a database as your data source, the configuration of the database affects the performance of accessing the data. By default, AutoAI opens 4 parallel connections to a database to speed up the data download. However, if the configuration of the database does not permit 4 connections, AutoAI rolls back to downloading by using 1 connection at a time. Configuring the database to accept more connections improves the data access performance.
• The maximum number of feature columns for a classification or regression experiment is 5000.

Understanding the AutoAI process

For additional detail on each of these phases, including links to associated research papers and descriptions of the algorithms applied to create the model pipelines, see AutoAI implementation details.

Data pre-processing

Most data sets contain different data formats and missing values, but standard machine learning algorithms work only with numbers and no missing values. Therefore, AutoAI applies various algorithms or estimators to analyze, clean, and prepare your raw data for machine learning. This technique automatically detects and categorizes values based on features, such as data type: categorical or numerical. Depending on the categorization, AutoAI uses hyper-parameter optimization to determine the best combination of strategies for missing value imputation, feature encoding, and feature scaling for your data.

Automated model selection

AutoAI uses automated model selection to identify the best model for your data. This novel approach tests potential models against small subsets of the data and ranks them based on accuracy. AutoAI then selects the most promising models and increases the size of the data subset until it identifies the best match. This approach saves time and improves performance by gradually narrowing down the potential models based on accuracy.

For information on how to handle automatically-generated pipelines to select the best model, refer to Selecting an AutoAI model.

Automated feature engineering

Feature engineering identifies the most accurate model by transforming raw data into a combination of features that best represent the problem. This unique approach explores various feature construction choices in a structured, nonexhaustive manner, while progressively maximizing model accuracy by using reinforcement learning. This technique results in an optimized sequence of transformations for the data that best match the algorithms of the model selection step.

Hyperparameter optimization

Hyperparameter optimization refines the best performing models. AutoAI uses a novel hyperparameter optimization algorithm for certain function evaluations, such as model training and scoring, that are typical in machine learning. This approach quickly identifies the best model despite long evaluation times at each iteration.

Ensembling and incremental learning

The process of building BatchedTreeEnsemble pipelines on top of the ranked pipelines. The ensemble pipelines provides incremental learning capabilities, and can be used to continue training by using the remaining data in a subsampled source, dividing the remaining data into batches, if needed. Each batch of training data is scored independently by using the optimized metric, so you can review the performance of each batch when you explore the results. For details, see Incremental learning.