classifiers.xml

使用具有增量学习的监控式学习方法。包括几个不同的分类算法。

<chapter id="classifiers">
<title>Classifiers</title>

<para>
This chapter describes (some of) the different classifiers available in <application>Select</application>.
</para>

<para>
There are two main types of classifiers in <application>Select</application>: document classifiers and vector classifiers.
</para>

<sect1>
<title>Document classifiers</title>
<para>
A document classifier is implemented as taking a document as input.
</para>

<para>
All document classifiers have type document.
<programlisting>type		document		# Document classifier
</programlisting>
</para>

<sect2>
<title>From classifier</title>
<para>
Classifies according to sender.
</para>

<sect3>
<title>Options</title>

<variablelist>
<varlistentry>
<term><option>n</option></term>
<listitem>
<simpara>
Specifies the maximum number of addresses to save.
</simpara>
</listitem>
</varlistentry>
<varlistentry>
<term><option>o</option></term>
<listitem>
<simpara>
Specifies the address eviction order.
</simpara>
</listitem>
</varlistentry>
</variablelist>

</sect3>

<sect3>
<title>Example</title>
<para>
<programlisting>[classifier]
name		from			# Name the classifier from
classifier	From			# From classifier
type		document		# Document classifier
options		n=100,o=fifo		# 100 addresses, fifo order
</programlisting>
</para>
</sect3>

</sect2>

<sect2>
<title>Reply classifier</title>
<para>
Classifies according to threads.
</para>

<sect3>
<title>options</title>

<variablelist>
<varlistentry>
<term><option>n</option></term>
<listitem>
<simpara>
Specifies the maximum number of threads to save.
</simpara>
</listitem>
</varlistentry>
</variablelist>

</sect3>

<sect3>
<title>Example</title>
<para>
<programlisting>[classifier]
name		reply			# Name the classifier reply
classifier	Reply			# Reply classifier
type		document		# Document classifier
options		n=100			# 100 subject entries
</programlisting>
</para>
</sect3>

</sect2>

</sect1>

<sect1>
<title>Vector classifiers</title>
<para>
A vector classifier is implemented as taking a vector as input.
</para>

<para>
Vector classifiers can have different type arguments.
A multi class classifier should have type multi_one:
<programlisting>type		multi_one		# Multi classifier, type ONE_MAX
</programlisting>
A binary classifier can have one of several types:
<programlisting>type		multi_rest		# Multi classifier, type REST_MAX
type		multi_linmax		# Multi classifier, type LIN_MAX
type		multi_uc		# Multi classifier, type UC_MAX
</programlisting>
</para>

<sect2>
<title>Alma</title>
<para>
Alma is a binary maximal margin classifier.
See <link linkend="Gen01"><citation>Gen01</citation></link> for a description of it.
</para>

<sect3>
<title>options</title>
<para>
There are no options.
</para>
</sect3>

<sect3>
<title>Example</title>
<para>
<programlisting>[classifier]
name		alma			# Name the classifier alma
classifier	Alma			# Alma classifier
type		multi_linmax		# Multi classifier, type LIN_MAX
options					# No options
tokenizer	alpha			# Alpha tokenizer
vectorizer	tfidf			# TF-IDF vectorizer
normalizer				# No normalization
</programlisting>
</para>
</sect3>

</sect2>

<sect2>
<title>Naive Bayes</title>
<para>
Naive Bayes is a simple probabilistic multi class classifier.
See <link linkend="McCNig98"><citation>McCNig98</citation></link> for a description of it.
</para>

<para>
Should only be used with type multi_one.
</para>

<sect3>
<title>options</title>
<para>
There are no options.
</para>
</sect3>

<sect3>
<title>Example</title>
<para>
<programlisting>[classifier]
name		nb			# Name the classifier nb
classifier	NaiveBayes		# NaiveBayes classifier
type		multi_one		# Multi classifier, type ONE_MAX
options					# No options
tokenizer	alpha			# Alpha tokenizer
vectorizer	tfidf			# TF-IDF vectorizer
normalizer				# No normalization
</programlisting>
</para>
</sect3>

</sect2>

<sect2>
<title>N-gram</title>
<para>
N-gram is a classifier which uses relative entropy.
It is suitable to use for language identification.
See <link linkend="SibRey96"><citation>SibRey96</citation></link> for a description of it.
</para>

<para>
Should only be used with an n-gram tokenizer and type multi_one.
</para>

<sect3>
<title>options</title>
<para>
There are no options.
</para>
</sect3>

<sect3>
<title>Example</title>
<para>
<programlisting>[classifier]
name		ng			# Name the classifier ng
classifier	N-gram			# N-gram classifier
type		multi_one		# Multi classifier, type ONE_MAX
options					# No options
tokenizer	ngram.byte		# N-gram byte tokenizer
vectorizer	tf			# TF vectorizer
normalizer				# No normalization
</programlisting>
</para>
</sect3>

</sect2>

<sect2>
<title>Perceptron</title>
<para>
Perceptron is an old, simple binary classifier.
It is described in just about every textbook on machine learning.
</para>

<para>
It can be used with type multi_rest, multi_linmax, multi_uc.
</para>

<sect3>
<title>options</title>
<para>
There are no options.
</para>
</sect3>

<sect3>
<title>Example</title>
<para>
<programlisting>[classifier]
name		per			# Name the classifier per
classifier	Perceptron		# Perceptron classifier
type		multi_linmax		# Multi classifier, type LIN_MAX
options					# No options
tokenizer	alpha			# Alpha tokenizer
vectorizer	tfidf			# TF-IDF vectorizer
normalizer				# No normalization
</programlisting>
</para>
</sect3>

</sect2>

<sect2 id="trivial">
<title>Trivial classifier</title>
<para>
Classifies classifies either according to class frequency or at random.
Needless to say, this is only useful for testing purposes and should not be used in practice.
</para>

<para>
Should only be used with a null tokenizer and type multi_one.
</para>

<sect3>
<title>Options</title>

<variablelist>
<varlistentry>
<term><option>s</option></term>
<listitem>
<simpara>
Seed (!= 0) for the pseudo random number generator. Sets the classifier to random mode.
</simpara>
</listitem>
</varlistentry>
</variablelist>

</sect3>

<sect3>
<title>Example</title>
<para>
<programlisting>[classifier]
name		triv			# Name the classifier triv
classifier	Trivial			# Trivial classifier
type		multi_one		# Multi classifier, type ONE_MAX
options		s=123			# Random mode, with seed 123
tokenizer	null			# Null tokenizer
vectorizer				# Default vectorizer
normalizer				# No normalization
</programlisting>
</para>
</sect3>

</sect2>

</sect1>

</chapter>