数学表达式: 从恐惧到单挑 (14. 决策表)

14. 决策表

决策表是机器学习中常用的概念.

14.1 数值数据

如果采集的数据均为实型值, 标签为二值的 (是/否, 0/1), 相应的决策表称为二分类决策系统.
Definition 1. A binary class decision system is a tuple $S=(\mathbf{X},\mathbf{Y})$ where $\mathbf{X}=[x_{ij}{]}_{n\times m}\in {\mathbb{R}}^{n\times m}$ is the data matrix, Y = [ y 1 , y 2 , … , y n ] ∈ { 0 , 1 } n \mathbf{Y} = [y_1, y_2, \dots, y_n] \in \{0, 1\}^n Y=[y1​,y2​,…,yn​]∈{0,1}n is the label array, $n$ is the number of instances, and $m$ is the number of features.
Naturally, $y_i$ is the label of instance ${\mathbf{x}}_i=[x_{i1},x_{i_2},\dots ,x_{im}]$.

• 在某些文献中, $m$ 和 $n$ 可能互换, 即每一列表示一个实例.
• 讨论: $n$, $m$ 是否应该出现在元组里面?

Java 代码:

public class BinaryDecisionSystem {
	/**
	 * 数据矩阵.
	 */
	double[][] data;

	/**
	 * 决策属性.
	 */
	boolean[] decision;

	/**
	 * 初始化一个决策系统.
	 */
	public BinaryDecisionSystem(int paraN, int paraM) {
		data = new double[paraN][paraM];
		decision = new boolean[paraN];
	}// Of the constructor
}// Of class BinaryDecisionSystem

以下各定义, 均建议读者先自己写一个版本, 再与我的对照.

如果标签为多值的 (高/中/低, 1/2/…/$d$). 相应的决策表称为多分类决策系统.

Definition 2. A multi-class decision system is a tuple $S=(\mathbf{X},\mathbf{Y})$ where $\mathbf{X}=[x_{ij}{]}_{n\times m}\in {\mathbb{R}}^{n\times m}$ is the data matrix, Y = [ y 1 , y 2 , … , y n ] ∈ { 1 , 2 , … , d } n \mathbf{Y} = [y_1, y_2, \dots, y_n] \in \{1, 2, \dots, d\}^n Y=[y1​,y2​,…,yn​]∈{1,2,…,d}n is the label array, $n$ is the number of instances, $m$ is the number of features, and $d$ is the number of classes.

如果有多个标签, 且均为二值, 相应的决策表称为多标签决策系统.
Definition 3. A multi-label decision system is a tuple $S=(\mathbf{X},\mathbf{Y})$ where $\mathbf{X}=[x_{ij}{]}_{n\times m}\in {\mathbb{R}}^{n\times m}$ is the data matrix, Y = [ y i k ] n × l ∈ { 0 , 1 } n × l \mathbf{Y} = [y_{ik}]_{n \times l} \in \{0, 1\}^{n \times l} Y=[yik​]n×l​∈{0,1}n×l is the label matrix, $n$ is the number of instances, $m$ is the number of features, and $l$ is the number of labels.

14.2 符号型数据

表 1. 符号型决策表.

Patient	Headache	Temperature	Lymphocyte	Leukocyte	Eosinophil	Heartbeat	Flu
$x_1$	Yes	High	High	High	High	Normal	Yes
$x_2$	Yes	High	Normal	High	High	Abnormal	Yes
$x_3$	Yes	High	High	High	Normal	Abnormal	Yes
$x_4$	No	High	Normal	Normal	Normal	Normal	No
$x_5$	Yes	Normal	Normal	Low	High	Abnormal	No
$x_6$	Yes	Normal	Low	High	Normal	Abnormal	No
$x_7$	Yes	Low	Low	High	Normal	Normal	Yes

思考:

• 属性值为符号型, 怎么办?
• 不同属性的定义域不同, 怎么办?

Definition 4. A decision system is a 5-tuple $S=(\mathbf{U},\mathbf{C},\mathbf{D},\mathbf{V},I)$, where

• U = { x 1 , x 2 , … , x n } \mathbf{U} = \{x_1, x_2, \dots, x_n\} U={x1​,x2​,…,xn​} is the set of instances,
• C = { a 1 , a 2 , … , a m } \mathbf{C} = \{a_1, a_2, \dots, a_m\} C={a1​,a2​,…,am​} is the set of conditions attributes,
• D = { d 1 , d 2 , … , d l } \mathbf{D} = \{d_1, d_2, \dots, d_l\} D={d1​,d2​,…,dl​} is the set of decisional attributes,
• $\mathbf{V}={\bigcup }_{a\in \mathbf{C}\cup \mathbf{D}}{\mathbf{V}}_a$,
• ${\mathbf{V}}_a$ is the domain of $a\in \mathbf{C}\cup \mathbf{D}$,
• $I:\mathbf{U}\times (\mathbf{C}\cup \mathbf{D})\to \mathbf{V}$ is the information function.

说明:

• $\mathbf{U}$, $\mathbf{C}$, 和 $\mathbf{D}$ 本质上是枚举型的;
• 使用 $\mathbf{V}$ 是为了方便;
• $\mathbf{C}$ 和 $\mathbf{D}$ 应分开;
• $I$ 是映射 (函数).

课堂练习:

• 写出本例中的 $\mathbf{U}$, $\mathbf{C}$, $\mathbf{D}$ 和 $\mathbf{V}$. 注: 最后两个属性为决策属性.
• $I$ 是怎么表示的?

Java 代码:

public enum ValuesEnum {
	YES, NO, HIGH, LOW, NORMAL, ABNORMAL;
}//Of enum ValuesEnum

public class DecisionSystem {
	/**
	 * Number of instances.
	 */
	int numInstances;

	/**
	 * Number of conditions.
	 */
	int numConditions;

	/**
	 * Number of decisions.
	 */
	int numDecisions;

	/**
	 * The data.
	 */
	ValuesEnum data[][];

	/**
	 * Construct the data.
	 */
	public DecisionSystem(int paraN, int paraM, int paraL) {
		numInstances = paraN;
		numConditions = paraM;
		numDecisions = paraL;
		data = new ValuesEnum[numInstances][numConditions + numDecisions];
	}// Of the first constructor
}// Of class DecisionSystem

进一步思考: 可否将属性写成函数?
Definition 5. A decision system is a quadruple $S=(\mathbf{U},\mathbf{C},\mathbf{D},\mathbf{V})$, where

• U = { x 1 , x 2 , … , x n } \mathbf{U} = \{x_1, x_2, \dots, x_n\} U={x1​,x2​,…,xn​} is the set of instances,
• C = { a 1 , a 2 , … , a m } \mathbf{C} = \{a_1, a_2, \dots, a_m\} C={a1​,a2​,…,am​} is the set of conditions attributes,
• D = { d 1 , d 2 , … , d l } \mathbf{D} = \{d_1, d_2, \dots, d_l\} D={d1​,d2​,…,dl​} is the set of decisional attributes,
• $\forall a\in \mathbf{C}\cup \mathbf{D}$, $a:\mathbf{U}\to {\mathbf{V}}_{\mathbf{a}}$ is an information function,
• $\mathbf{V}={\bigcup }_{a\in \mathbf{C}\cup \mathbf{D}}{\mathbf{V}}_a$.

讨论:

• 每个属性都是一个函数, 函数是否可以作为集合的元素?
• 是否可以简写为 $S=(\mathbf{U},\mathbf{C},\mathbf{D})$?
• 是否可以简写为 $S=(\mathbf{U},\mathbf{C}\cup \mathbf{D})$？

14.4 决策系统体系

结合论文 A hierarchical model for test-cost-sensitive decision systems 讲解.

14.5 作业

• 定义一个标签分布系统, 即各标签的值不是 0/1, 而是 $[0,1]$ 区间的实数, 且同一对象的标签和为 1.