demo 代码:传送门
引言
上次分享Apriori算法时,我们有提到Apriori算法在每次增加频繁项集的大小时,会重新扫描整个数据集。当数据集很大时,这会显著降低频繁项集发现的速度。而本次分享的FP-growth(frequent patten)算法就能高效地发现频繁项集。
那么在现实生活中,是否存在应用FP-growth算法的产品呢?答案是存在的,如下图所示:
上图中,我们在Google搜索引擎里输入why这个单词,搜索引擎会自动给我们一些推荐查询词项,这些查询词项来自互联网上经常出现的词对。查找的方法正是接下来我们要分享的FP-growth算法。
##FP-growth与Apriori的区别
我们先从宏观上来看看两个算法的区别:
FP-growth算法只需要对数据库进行两次扫描,而Apriori算法对每个潜在的频繁项集都会扫描数据集判定给定模式是否频繁,因此FP-growth算法的速度要比Apriori算法快。在小规模数据集上,这不是什么问题,但当处理更大数据集时,FP-growth通常性能要比Apriori好两个数量级以上。
Apriori是发现频繁项集与关键规则的算法,但是FP-growth并不能用于发现关联规则。
##FP-growth基本流程
FP-growth只会扫描数据集两次,它在发现频繁项集的基本过程如下:
1.构建FP树
2.从FP树中挖掘频繁项集
FP树
我们先来看个例子,下图有一份数据样例:
移除非频繁项后的数据样例:
这份数据左侧代表事务ID,右侧代表事务中的元素项。那么根据上述数据样例生成的FP树如下所示:
与搜索树不同的是,一个元素项可以在一棵FP树中出现多次。其中树节点给出的是集合中单个元素及其在序列中的出现次数,路径会给出该序列出现的次数。带箭头的链接叫做节点链接,表示相似项之间的链接。
构建FP树的工作流程
那么我们来详细说下该算法的工作流程。
为构建FP树,需要对原始数据集扫描两遍。第一遍对所有元素项的出现次数进行计数。并应用上次分享的Apriori原理,如果某个元素是不频繁的,那么包含该元素的超集也是不频繁的,所以不需要考虑这些超集。因此数据集的第一次扫描是用来统计元素出现的频率的。第二次扫描就只考虑剩余的频繁元素。
因此FP树构建部分流程如下所示:
构建FP树
第二次扫描FP树就需要构建FP树了。因为FP树比较复杂,因此创建一个类来实现保存树的每个节点。
代码如下:
class treeNode:
def __init__(self, nameValue, numOccur, parentNode):
self.name = nameValue
self.count = numOccur
self.nodeLink = None
self.parent = parentNode #needs to be updated
self.children = {}
def inc(self, numOccur):
self.count += numOccur
def disp(self, ind=1):
print(' '*ind, self.name, ' ', self.count)
for child in self.children.values():
child.disp(ind+1)
上面的程序给出了FP树种节点的类的定义。
除了上图FP树之外,还需要建立一个头指针表用来快速访问FP树中一个给定类型的所有元素。如下图所示:
上图中的Header Table即头指针表,这里使用一个字典来作为数据结构存储头指针表。
接下来第二次扫描数据集并构建FP树之前,需要将每个集合的元素进行排序,从而保证相同项只会表现一次。
FP构建函数如下所示:
def createTree(dataSet, minSup=1): #create FP-tree from dataset but don't mine
headerTable = {}
#go over dataSet twice
for trans in dataSet:#first pass counts frequency of occurance
for item in trans:
headerTable[item] = headerTable.get(item, 0) + dataSet[trans]
for k in headerTable.keys(): #remove items not meeting minSup
if headerTable[k] < minSup:
del(headerTable[k])
freqItemSet = set(headerTable.keys())
#print 'freqItemSet: ',freqItemSet
if len(freqItemSet) == 0: return None, None #if no items meet min support -->get out
for k in headerTable:
headerTable[k] = [headerTable[k], None] #reformat headerTable to use Node link
#print 'headerTable: ',headerTable
retTree = treeNode('Null Set', 1, None) #create tree
for tranSet, count in dataSet.items(): #go through dataset 2nd time
localD = {}
for item in tranSet: #put transaction items in order
if item in freqItemSet:
localD[item] = headerTable[item][0]
if len(localD) > 0:
orderedItems = [v[0] for v in sorted(localD.items(), key=lambda p: p[1], reverse=True)]
updateTree(orderedItems, retTree, headerTable, count)#populate tree with ordered freq itemset
return retTree, headerTable #return tree and header table
def updateTree(items, inTree, headerTable, count):
if items[0] in inTree.children:#check if orderedItems[0] in retTree.children
inTree.children[items[0]].inc(count) #incrament count
else: #add items[0] to inTree.children
inTree.children[items[0]] = treeNode(items[0], count, inTree)
if headerTable[items[0]][1] == None: #update header table
headerTable[items[0]][1] = inTree.children[items[0]]
else:
updateHeader(headerTable[items[0]][1], inTree.children[items[0]])
if len(items) > 1:#call updateTree() with remaining ordered items
updateTree(items[1::], inTree.children[items[0]], headerTable, count)
def updateHeader(nodeToTest, targetNode): #this version does not use recursion
while (nodeToTest.nodeLink != None): #Do not use recursion to traverse a linked list!
nodeToTest = nodeToTest.nodeLink
nodeToTest.nodeLink = targetNode
上述代码中包含三个函数。第一个函数createTree()是用来构建FP树。可以从代码中看到两个嵌套for循环代表遍历数据集两次。updateTree()及updateHeader()都是用来更新FP树及头指针表的。
实验结果如下:
simpDat = fpGrowth.loadSimpDat()
simpDat
[['r', 'z', 'h', 'j', 'p'],
['z', 'y', 'x', 'w', 'v', 'u', 't', 's'],
['z'],
['r', 'x', 'n', 'o', 's'],
['y', 'r', 'x', 'z', 'q', 't', 'p'],
['y', 'z', 'x', 'e', 'q', 's', 't', 'm']]
initSet = fpGrowth.createInitSet(simpDat)
initSet
{frozenset({'z'}): 1,
frozenset({'h', 'j', 'p', 'r', 'z'}): 1,
frozenset({'s', 't', 'u', 'v', 'w', 'x', 'y', 'z'}): 1,
frozenset({'n', 'o', 'r', 's', 'x'}): 1,
frozenset({'p', 'q', 'r', 't', 'x', 'y', 'z'}): 1,
frozenset({'e', 'm', 'q', 's', 't', 'x', 'y', 'z'}): 1}
importlib.reload(fpGrowth)
myFPtree,myHeaderTab = fpGrowth.createTree(initSet,3)
myFPtree.disp()
Null Set 1
z 5
r 1
x 3
s 2
t 2
y 2
t 1
r 1
y 1
x 1
s 1
r 1
从一棵FP树中挖掘出频繁项集
这里挖掘频繁项集的思路与Apriori算法类似,首先从单元素项开始,然后在此基础上逐步构建更大的集合。因为整个挖掘频繁项集的过程是围绕FP树进行的,因此不再需要原始数据集了。
从FP树中抽取频繁项集的三个基本步骤如下:
- 从FP树中获得条件模式基;
- 利用条件模式基,构建一个条件FP树;
- 迭代重复步骤(1)步骤(2),直到树包含一个元素项为止。
####条件模式基
条件模式基是以所查找元素项为结尾的路径集合。每一条路径其实都是一条前缀路径。下图给出了每个频繁项的所有前缀路径。
从上图我们可以发现获得前缀路径的方式:利用先前创建的头指针,来遍历并回溯。
条件FP树
对于每一个频繁项,都要创建一棵条件FP树。然后在构建条件FP的过程中,再将不满足最小支持度的元素项去掉。具体代码如下:
def mineTree(inTree, headerTable, minSup, preFix, freqItemList):
bigL = [v[0] for v in sorted(headerTable.items(), key=lambda p: p[1])]#(sort header table)
for basePat in bigL: #start from bottom of header table
newFreqSet = preFix.copy()
newFreqSet.add(basePat)
#print 'finalFrequent Item: ',newFreqSet #append to set
freqItemList.append(newFreqSet)
condPattBases = findPrefixPath(basePat, headerTable[basePat][1])
#print 'condPattBases :',basePat, condPattBases
#2. construct cond FP-tree from cond. pattern base
myCondTree, myHead = createTree(condPattBases, minSup)
#print 'head from conditional tree: ', myHead
if myHead != None: #3. mine cond. FP-tree
#print 'conditional tree for: ',newFreqSet
#myCondTree.disp(1)
mineTree(myCondTree, myHead, minSup, newFreqSet, freqItemList)
运行效果:
importlib.reload(fpGrowth)
myFPtree,myHeaderTab = fpGrowth.createTree(initSet,3)
myFPtree.disp()
Null Set 1
z 5
r 1
x 3
s 2
t 2
y 2
t 1
r 1
y 1
x 1
s 1
r 1
importlib.reload(fpGrowth)
freqItems = []
fpGrowth.mineTree(myFPtree,myHeaderTab,3,set([]),freqItems)
finalFrequent Item: {'r'}
condPattBases : r {frozenset({'z'}): 1, frozenset({'x', 's'}): 1, frozenset({'t', 'x', 'z'}): 1}
head from conditional tree: None
finalFrequent Item: {'s'}
condPattBases : s {frozenset({'x', 'z'}): 2, frozenset({'x'}): 1}
head from conditional tree: {'x': [3, <fpGrowth.treeNode object at 0x000001DEEEFA65C0>]}
conditional tree for: {'s'}
Null Set 1
x 3
finalFrequent Item: {'x', 's'}
condPattBases : x {}
head from conditional tree: None
finalFrequent Item: {'t'}
condPattBases : t {frozenset({'x', 'z', 's'}): 2, frozenset({'x', 'z'}): 1}
head from conditional tree: {'x': [3, <fpGrowth.treeNode object at 0x000001DEEEFA0240>], 'z': [3, <fpGrowth.treeNode object at 0x000001DEEEFA0278>]}
conditional tree for: {'t'}
Null Set 1
x 3
z 3
finalFrequent Item: {'t', 'x'}
condPattBases : x {}
head from conditional tree: None
finalFrequent Item: {'t', 'z'}
condPattBases : z {frozenset({'x'}): 3}
head from conditional tree: {'x': [3, <fpGrowth.treeNode object at 0x000001DEEEFA5128>]}
conditional tree for: {'t', 'z'}
Null Set 1
x 3
finalFrequent Item: {'t', 'x', 'z'}
condPattBases : x {}
head from conditional tree: None
finalFrequent Item: {'x'}
condPattBases : x {frozenset({'z'}): 3}
head from conditional tree: {'z': [3, <fpGrowth.treeNode object at 0x000001DEEEFA5358>]}
conditional tree for: {'x'}
Null Set 1
z 3
finalFrequent Item: {'x', 'z'}
condPattBases : z {}
head from conditional tree: None
finalFrequent Item: {'y'}
condPattBases : y {frozenset({'t', 'x', 'z', 's'}): 2, frozenset({'t', 'x', 'z', 'r'}): 1}
head from conditional tree: {'t': [3, <fpGrowth.treeNode object at 0x000001DEEEFA5F28>], 'x': [3, <fpGrowth.treeNode object at 0x000001DEEEFA5F60>], 'z': [3, <fpGrowth.treeNode object at 0x000001DEEEFA5F98>]}
conditional tree for: {'y'}
Null Set 1
t 3
x 3
z 3
finalFrequent Item: {'t', 'y'}
condPattBases : t {}
head from conditional tree: None
finalFrequent Item: {'y', 'x'}
condPattBases : x {frozenset({'t'}): 3}
head from conditional tree: {'t': [3, <fpGrowth.treeNode object at 0x000001DEEEFB4080>]}
conditional tree for: {'y', 'x'}
Null Set 1
t 3
finalFrequent Item: {'t', 'y', 'x'}
condPattBases : t {}
head from conditional tree: None
finalFrequent Item: {'y', 'z'}
condPattBases : z {frozenset({'t', 'x'}): 3}
head from conditional tree: {'t': [3, <fpGrowth.treeNode object at 0x000001DEEEFB4C50>], 'x': [3, <fpGrowth.treeNode object at 0x000001DEEEFB4C88>]}
conditional tree for: {'y', 'z'}
Null Set 1
t 3
x 3
finalFrequent Item: {'t', 'y', 'z'}
condPattBases : t {}
head from conditional tree: None
finalFrequent Item: {'y', 'z', 'x'}
condPattBases : x {frozenset({'t'}): 3}
head from conditional tree: {'t': [3, <fpGrowth.treeNode object at 0x000001DEEEF8C860>]}
conditional tree for: {'y', 'z', 'x'}
Null Set 1
t 3
finalFrequent Item: {'t', 'y', 'z', 'x'}
condPattBases : t {}
head from conditional tree: None
finalFrequent Item: {'z'}
condPattBases : z {}
head from conditional tree: None
总结
FP-growth算法是一种发现数据集中频繁模式的有效方法。可以使用FP-growth在多种文档中查找频繁单词。频繁项集的生成有着许多应用,比如购物交易、医学诊断及大气研究等。
标签:FP,condPattBases,items,conditional,tree,算法,growth,headerTable From: https://blog.51cto.com/u_15996214/6105908