SAS^® High-Performance Analytics

Features

SAS^® High-Performance Statistics

High-performance linear regression

• Supports generalized linear models and reference parameterization for classification effects.
• Supports partitioning of data into training, validation and testing roles.
• Supports a FREQ statement for grouped analysis and a WEIGHT statement for weighted analysis.
• Provides multiple effect-selection methods.

High-performance nonlinear regression

• Computes analytical derivatives of user-provided expressions for more robust parameter estimations, both improving and accelerating the estimations.
• Evaluates user-provided expressions and their confidence limits with the ESTIMATE and PREDICT statements.
• Estimates parameters by using least squares and the maximum likelihood method.

High-performance logistic regression

• Predicts binary, binomial and multinomial outcomes.
• Provides model-building syntax with the CLASS and effect-based MODEL statements (in particular, the GLM, LOGISTIC, GLIMMIX and MIXED procedures).
• Provides cumulative link models for ordinal data; generalized logistic modeling for unordered multinomial data enables model building (variable selection) through the SELECTION statement.
• Provides a WEIGHT statement for weighted analysis and a FREQ statement for grouped analysis.
• Provides an OUTPUT statement to produce a data set with predicted probabilities and other observation-wise statistics.

High-performance mixed linear models

• Fits a variety of mixed linear models to data and enables you to use these fitted models to make statistical inferences about the data.
• Supports multiple covariance structures.
• Provides appropriate standard errors for all specified, estimable linear combinations of fixed and random effects, and corresponding t-tests and F-tests.
• REML and ML estimation methods are implemented with a variety of optimization algorithms.
• Provides special dense and sparse algorithms that take advantage of distributed and multiple-core computing environments.

SAS^® High-Performance Data Mining

High-performance neural networks

• Provides automatic standardization of input and target variables.
• Provides intelligent defaults for most neural network parameters such as activation and error functions.
• Provides automatic selection and use of a validation data subset.
• Provides automatic termination of training when the validation error stops improving.
• Provides the ability to weight individual observations.

High-performance forests

• Creates an ensemble of hundreds of decision trees to predict a single target.
• Trains hundreds of decision trees in parallel independently on different grid nodes.
• Randomly selects the input variables considered for splitting a node from all available inputs.
• Considers only a single variable that is most associated with the target for splitting.

High-performance 4SCORE

• The 4SCORE procedure scores a previously trained forest model produced by the HPFOREST procedure.

High-performance data mining nodes

• Includes the following high-performance-enabled SAS Enterprise Miner nodes:

  • HP Data Partition.
  • HP Explore.
  • HP Transform.
  • HP Variable Selection.
  • HP Regression.
  • HP Neural.
  • HP Forest.
  • HP Impute.

High-derformance decide

• Creates optimal decisions that are based on a decision matrix that you specify, on prior probabilities, and on output from a modeling procedure (which can be predicted values for an interval target variable).
• The decision matrix contains columns (decision variables) that correspond to each decision and rows (observations) that correspond to target values. The values of the decision variables represent target-specific consequences, which might be profit, loss, or revenue.

High-performance variable reduction

• Reduces dimensionality by using the HPREDUCE procedure for structured inputs and to select a subset of the original variables (variable selection) to preserve model interpretation.
• Performs unsupervised variable section by identifying a set of variables that jointly explain the maximum amount of data variance (covariance analysis).
• Provides distributed computation and output of the CORR, COV or SSCP matrix.
• Uses the CLASS statement to support categorical inputs.
• Supports main and interaction effects with the VAR statement.
• Outputs statistics and matrix information for exploratory data analysis that can also be used as direct input for statistical procedures, thus saving time by eliminating redundant matrix aggregations.

SAS^® High-Performance Econometrics

High-performance count regression

• Fits regression models where the dependent variable represents counts (e.g., the number of events recorded in some period or for some subject).
• Supports Poisson and negative binomial models.
• Supports zero-inflated Poisson and negative binomial models, and can fit separate regressors for the zero-inflated distribution.
• Estimates parameters by using the maximum likelihood method.

High-performance severity models

• Fits probability distributions for the severity (magnitude) of random events (including those with negative effects – e.g., the magnitude of damages caused by natural disasters, distributions of losses claimed under insurance policies, or the severity of disease outbreaks – as well as events with positive effects – e.g., the intermittent demand for certain products).
• Fits regression models for the scale of the severity distribution.
• Provides nine different probability distributions, including the Tweedie distribution, and can automatically select the best-fitting distribution.
• Allows users to add additional probability distributions.
• Can model data truncation and data censoring.

High-performance qualitative and limited independent variable models

• Fits linear, censored and truncated regression models with heteroscedasticity and stochastic frontier production and cost models.

SAS^® High-Performance Text Mining

• Text parsing with properties that include:
  • Detection of different parts of speech, stemming and synonyms.
  • Frequency and term weighting.
• Transformation via singular value decomposition (SVD) that:
  • Reduces the term-by-document matrix generated from the parsing process to a reduced numeric, structured representation of the document collection.
  • Enables transformation output to be used as input into high-performance structured data mining procedures.
• Graphs and table outputs provide detailed information about the terms and their distributions within the collection.
• Includes high-performance text mining scoring of large-scale textual data.

SAS^® High-Performance Optimization

High-performance local search optimization

• Ability to optimize a user-defined objective subject to nonlinear constraints.
• Allows continuous and integer variables.
• Ability to embed the genetic algorithm solver (PROC GA) from SAS/OR, as well as other local search techniques.
• Available only for Teradata and Greenplum at this time.

High-performance optimization

• Provides a decomposition algorithm that works well for certain classes of large structured linear and mixed integer optimization problems. The algorithm decomposes the overall problem into a set of component problems that can be solved quickly.
• Provides a multistart capability that increases the likelihood of identifying a globally optimal solution. It can be used to select and begin optimization from each point. The best solution found among all starting points is reported.
• Parallelized tuner functionality in PROC OPTMILP helps define a good option setting for certain problems. It solves lots of problems with various settings using a clever search of the option combinations and finds the best set of option values.
• Available only for Teradata and Greenplum at this time.

SAS^® High-Performance Forecasting

• Designed to efficiently process large-scale, hierarchically-structured, time-stamped data sets. Process multiple hierarchy levels in a single pass of the data set.
• Provides an automatic way to generate time series from the time-stamped data, and produce forecasts for them in one step.
• For typical time series, the HPFORECAST procedure automatically chooses the best performing forecast from among the following smoothing models: Simple, linear, damped trend, seasonal (additive and multiplicative), Winters method (additive and multiplicative).
• Transformed versions of the models include: Log, square root, logistic, and Box-Cox.
• Can forecast time series data, whose observations are equally spaced by a specific time interval (for example, monthly, weekly), and also transactional data, whose observations are not spaced with respect to any particular time interval.

Common set of high-performance procedures

The following procedures are available in all six of the SAS High-Performance Analytics products listed above.

High-performance data summarization

• Enables large-scale data exploration and summarization through a series of parallelized procedures.
• Generates descriptive statistics on a large scale, very quickly.
• Creates mean, min, max, range and measures of spread and centrality along with data for cardinality, summary and levels of variables.

High-performance data mining database

• Creates summary statistics of key input data sources, using:

  • Number of observations.
  • Number of observations that contain a missing value.
  • Minimum observed value.
  • Maximum observed value.
  • Mean of observed values.
  • Standard deviation.
  • Measure of asymmetry.
  • Measure of the "heaviness of the tails."
  • Sum of all non-missing observations.
  • Corrected sum of squares.
  • Sum of squares.

High-performance correlation

• Enables you to compute correlations for big data sets that have large quantities of rows and columns.

High-performance sampling

• Performs either high performance simple random sampling or stratified sampling.
• Creates one output data set, which contains the sample data set; one performance table, which contains performance information; and one frequency table, which contains the frequency information for the population and sample.

High-performance binning

• Bucket (equal-length) binning method.
• Winsorized binning method and Winsorized statistics.
• Pseudo–quantile binning method, which is similar to quantile binning.
• Provides a mapping table for the selected binning method.
• Provides a basic statistical table that contains the minimum, maximum, mean, pseudo-median, and so on.
• Histogram table that shows the output mapping statistics.

High-performance imputation

• Executes high-performance numeric variable imputation with a specified value.
• Can also replace numeric missing values with the mean, the pseudo-median, or some random value between the minimum value and the maximum value of the non-missing values.