马尔可夫毯式遗传算法在基因选择中的应用

引用

LaTex

@article{ZHU20073236,
title = “Markov blanket-embedded genetic algorithm for gene selection”,
journal = “Pattern Recognition”,
volume = “40”,
number = “11”,
pages = “3236 - 3248”,
year = “2007”,
issn = “0031-3203”,
doi = “https://doi.org/10.1016/j.patcog.2007.02.007“,
url = “http://www.sciencedirect.com/science/article/pii/S0031320307000945“,
author = “Zexuan Zhu and Yew-Soon Ong and Manoranjan Dash”,
keywords = “Microarray, Feature selection, Markov blanket, Genetic algorithm (GA), Memetic algorithm (MA)”
}

Normal

Zexuan Zhu, Yew-Soon Ong, Manoranjan Dash,
Markov blanket-embedded genetic algorithm for gene selection,
Pattern Recognition,
Volume 40, Issue 11,
2007,
Pages 3236-3248,
ISSN 0031-3203,
https://doi.org/10.1016/j.patcog.2007.02.007.
(http://www.sciencedirect.com/science/article/pii/S0031320307000945)
Keywords: Microarray; Feature selection; Markov blanket; Genetic algorithm (GA); Memetic algorithm (MA)

---

摘要

Microarray technologies
the smallest possible set of genes

Markov blanket-embedded genetic algorithm (MBEGA) for gene selection problem

Markov blanket and predictive power in classifier model

filter, wrapper, and standard GA

evaluation criteria:
classification accuracy, number of selected genes, computational cost, and robustness

---

主要内容

---

Markov Blanket（Markov毯）

$F$ — 所有特征的集合
$C$ — 类别

一个特征$F_i$的Markov毯 定义如下：

定义（Markov毯）
$M$ — 一个特征子集（不包含$F_i$）
即，$M \in F$且$F_i \notin M$。
$M$为$F_i$的一个Markov毯，若
给定$M$，$F_i$是对于 $\left( F \cup C \right) - M - \left\{ F_i \right\}$条件独立的，
即，$P \left( F - M - \left\{ F_i \right\}, C | F_i, M \right) = P \left( F - M - \left\{ F_i \right\}, C | M \right)$

给定X，两个属性A与B是条件独立的，若$P \left( A | X, B \right) = P \left( A | X \right)$，也就是说，B并不能在X之外提供关于A的信息。若一个特征$F_i$在当前选择的特征子集中有一个Markov毯$M$，那么$F_i$在$M$之外关于$C$不能提供其他选择的特征的信息，因此，$F_i$能够安全移除。然而，决定特征的条件独立的计算复杂度通常非常高，因此，只使用一个特征来估计$F_i$的Markov毯。

定义（近似Markov毯）
对于两个特征$F_i$与$F_j$ $i\neq j$，$F_j$可看作为$F_i$的近似Markov毯，若$SU_{j,C} \geq SU_{i,C}$ 且 $SU_{i,j} \geq SU_{i,C}$，其中，
对称不确定性（symmetrical uncertainty，SU）度量特征（包括类，$C$）间的相关性，定义为：

$IG \left( F_i | F_j \right)$ — 特征$F_i$与$F_j$间的信息增益
$H \left( F_i \right)$与$H \left( F_j \right)$ — 特征$F_i$与$F_j$的熵
$SU_{i,C}$ — 特征$F_i$与类$C$间的相关性，称为C-correlation
一个特征被认为是相关的若其C-correlation高于用户给定的阈值$\gamma$，即，$S_{i,C} > \gamma$
没有任何近似Markov毯的特征为predominant feature主导特征

---

马尔可夫毯式嵌入式遗传算法

若适应值差异小于$\varepsilon$，则特征数较少的个体较好

Lamarckian learning：
通过将局部改进的个体放回种群竞争繁殖的机会，来迫使基因型反映改进的效果

$X$ — 选择的特征子集
$Y$ — 排除的特征子集

C-correlation 只计算一次

搜索范围$L$ — 定义了$Add$与$Del$操作的最大数目 — $L^2$个操作组合
随机顺序 — 直到得到改进提升效果

Lamarckian learning process

之后是
usual evolutionary operations：
1. linear ranking selection
2. uniform crossover
3. mutation operators with elitism

---

试验

MBEGA method

考虑了：
1. the FCBF (fast correlation-based filter)
2. BIRS (best incremental ranked subset)
3. standard GA feature selection algorithms

FCBF —
a fast correlation based filter method
1. selecting a subset of relevant features whose C-correlation are larger than a given threshold $\gamma$
2. sorts the relevant features in descending order in terms of C-correlation
3. redundant features are eliminated one-by-one in a descending order

A feature is redundant 仅当 it has an approximate Markov blanket

predominant features with zero redundant features in terms of C-correlation

BIRS — a similar scheme as the FCBF
evaluates the goodness of features using a classifier

1. ranking the genes according to some measure of interest
2. sequentially selects the ranked features one-by-one based on their incremental usefulness

calls to the classifier as many times as the number of features

$BIRS_F$ or $BIRS_W$ — 基于 — C-correlation (i.e., symmetrical uncertainty between feature $F_i$ and the class $C$) or individual predictive power

$BIRS_F$ 耗时更少

---

synthetic data 合成数据

ten 10-fold crossvalidations
with C4.5 classifier

10 independent runs

The maximum number of selected features in each chromosome, m, is set to 50.

---

microarray data 微阵列数据

The .632+ bootstrap

K次重采样

the support vector machine (SVM) — microarray classification problems

one-versus-rest strategy — multi-class datasets

the linear kernel SVM