From Jstacs
Revision as of 00:46, 2 February 2012 by Grau (talk | contribs) (Created page with "== General structure of Jstacs == A coarse view on the structure of Jstacs is presented in Figure #fig:classes (link). Being a library for statistical analysis and classific...")
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigationJump to search

General structure of Jstacs

A coarse view on the structure of Jstacs is presented in Figure #fig:classes (link). Being a library for statistical analysis and classification of sequence data, Jstacs is organized around the abstract class AbstractClassifier, the interface StatisticalModel and its two sub-interfaces TrainableStatisticalModel, and DifferentiableStatisticalModel.

Part of the class structure of Jstacs. Interfaces are depicted in red, abstract classes in blue, concrete classes in green, and enums in orange. Continuous transitions represent inheritance, whereas arrows represent usage.

StatisticalModel s represent statistical models in general, which can compute the log-likelihood of a given input sequence and define prior densities on their parameters. The abstract implementation AbstractTrainSM of TrainableStatisticalModel is the base class of many generatively learned models such as Bayesian networks, hidden Markov models, or mixture models, and can be learned from a single input data set. In constrast, DifferentiableStatisticalModel s provide all facilities for numerical optimization of parameters, which is especially necessary for discriminative parameter learning. The abstract base class of all DifferentiableStatisticalModel implementations is AbstractDifferentiableStatisticalModel. Currently, DifferentiableStatisticalModel s include Bayesian networks, Markov models, a ZOOPS model, and mixture models.

AbstractClassifier defines the general properties of a classifier. Its sub-class AbstractScoreBasedClassifier adds additional methods for the classification of sequences based on a sequence and class-specific score. Two concrete sub-classes of AbstractScoreBasedClassifier are the TrainSMBasedClassifier, which works on TrainableStatisticalModel s, and the GenDisMixClassifier, which works on DifferentiableStatisticalModel.

AbstractClassifier s can be assessed either on dedicated training and test data sets or in ClassifierAssessment s like KFoldCrossValidation or RepeatedHoldOutExperiment. The performance measures used in such an assessment are collected in PerformanceMeasureParameterSet s containing sub-classes of the abstract class AbstractPerformanceMeasure.

A more detailed view on all of these classes will be given in the remainder of this cookbook.

About this cookbook

This document is not a cookbook in the sense of a collection of recipes. The intention of this cookbook is rather to learn how to cook with the ingredients and tools provided by Jstacs.

Nonetheless, we present a collection of recipes in the last section, where you find the code of executable code examples that can also be downloaded from [xxx xxx].

We are aware that a library of this size seems daunting on first sight -- and on second sight as well. However, we hope that despite the inevitable complexity and size of such a library, this cookbook may help to get a picture of the structure, design principles, and capabilities of Jstacs.

This cookbook is structured as follows: in the first section, we explain how data are represented in Jstacs, and how you can read data from files.

In the second section, we present some facilities, an XML parser and the representation of parameters and results, that are used frequently within Jstacs and are necessary for the following parts.

In the third section, we present sequence scores, statistical models, and their sub-interfaces and sub-classes. We explain the methods defined in these interfaces and classes, and we show how you can create and use their existing implementations.

In the fourth section, we explain classifiers and assessment of classifiers using different performance measures.

In the fifth section, we present the facilities for numerical optimization.

In the sixth section, we list utility classes and methods, that we think might be of help for your own implementations.

Finally, in the last section, we give a number of executable code examples that may serve as a starting point of your own classes and applications.