广告系统中weak-and算法原理及编码验证

时间:2023-01-08 12:42:10

wand(weak and)算法基本思路

  一般搜索的query比较短,但如果query比较长,如是一段文本,需要搜索相似的文本,这时候一般就需要wand算法,该算法在广告系统中有比较成熟的应

该,主要是adsense场景,需要搜索一个页面内容的相似广告。

   Wand方法简单来说,一般我们在计算文本相关性的时候,会通过倒排索引的方式进行查询,通过倒排索引已经要比全量遍历节约大量时间,但是有时候仍

然很慢。

   原因是很多时候我们其实只是想要top n个结果,一些结果明显较差的也进行了复杂的相关性计算,而weak-and算法通过计算每个词的贡献上限来估计文档

的相关性上限,从而建立一个阈值对倒排中的结果进行减枝,从而得到提速的效果。

   wand算法首先要估计每个词对相关性贡献的上限,最简单的相关性就是TF*IDF,一般query中词的TF均为1,IDF是固定的,因此就是估计一个词在文档中的

词频TF上限,一般TF需要归一化,即除以文档所有词的个数,因此,就是要估算一个词在文档中所能占到的最大比例,这个线下计算即可。

   知道了一个词的相关性上界值,就可以知道一个query和一个文档的相关性上限值,显然就是他们共同的词的相关性上限值的和。

   这样对于一个query,获得其所有词的相关性贡献上限,然后对一个文档,看其和query中都出现的词,然后求这些词的贡献和即可,然后和一个预设值比

较,如果超过预设值,则进入下一步的计算,否则则丢弃。

  如果按照这样的方法计算n个最相似文档,就要取出所有的文档,每个文档作预计算,比较threshold,然后决定是否在top-n之列。这样计算当然可行,但

是还是可以优化的。

wand(weak and)算法原理演示

代码实现了主要的算法逻辑以验证算法的有效性,供大家参考,该实现优化了原始算法的一些逻辑尽量减少了无谓的循环:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
www.169it.com
#!/usr/bin/python
#wangben updated 20130108
class WAND:
    '''implement wand algorithm'''
    def __init__(self, InvertIndex, last_docid):
        self.invert_index = InvertIndex #InvertIndex: term -> docid1, docid2, docid3 ...
        self.current_doc = 0
        self.current_invert_index = {}
        self.query_terms = []
        self.threshold = 2
        self.sort_terms = []
        self.LastID = 2000000000 #big num
        self.debug_count = 0
        self.last_docid = last_docid
    def __InitQuery(self, query_terms):
        '''check terms len > 0'''
        self.current_doc = -1
        self.current_invert_index.clear()
        self.query_terms = query_terms
        self.sort_terms[:] = []
        self.debug_count = 0
        for term in query_terms:
            #initial start pos from the first position of term's invert_index
            self.current_invert_index[term] = [ self.invert_index[term][0], 0 ] #[ docid, index ]
                                                                                    
    def __SortTerms(self):
        if len(self.sort_terms) == 0:
            for term in self.query_terms:
                if term in self.current_invert_index:
                    doc_id = self.current_invert_index[term][0]
                    self.sort_terms.append([ int(doc_id), term ])
        self.sort_terms.sort()
                                                                                            
    def __PickTerm(self, pivot_index):
        return 0
    def __FindPivotTerm(self):
        score = 0
        for i in range(0, len(self.sort_terms)):
            score += 1
            if score >= self.threshold:
                return [ self.sort_terms[i][1], i]
        return [ None, len(self.sort_terms) ]
    def __IteratorInvertIndex(self, change_term, docid, pos):
        '''move to doc id > docid'''
        doc_list = self.invert_index[change_term]
        i = 0
        for i in range(pos, len(doc_list)):
            if doc_list[i] >= docid:
                pos = i
                docid = doc_list[i]
                break
        return [ docid, pos ]
    def __AdvanceTerm(self, change_index, docid ):
        change_term = self.sort_terms[change_index][1]
        pos = self.current_invert_index[change_term][1]
        (new_doc, new_pos) = \
            self.__IteratorInvertIndex(change_term, docid, pos)
                                                                                        
        self.current_invert_index[change_term] = \
            [ new_doc , new_pos ]
        self.sort_terms[change_index][0] = new_doc
                                                                                        
                                                                                        
    def __Next(self):
        if self.last_docid == self.current_doc:
            return None
                                                                                            
        while True:
            self.debug_count += 1
            #sort terms by doc id
            self.__SortTerms()
                                                                                            
            #find pivot term > threshold
            (pivot_term, pivot_index) = self.__FindPivotTerm()
            if pivot_term == None:
                #no more candidate
                return None
                                                                                            
            #debug_info:
            for i in range(0, pivot_index + 1):
                print self.sort_terms[i][0],self.sort_terms[i][1],"|",
            print ""
                                                                                                
            pivot_doc_id = self.current_invert_index[pivot_term][0]
            if pivot_doc_id == self.LastID: #!!
                return None
            if pivot_doc_id <= self.current_doc:
                change_index = self.__PickTerm(pivot_index)
                self.__AdvanceTerm( change_index, self.current_doc + 1 )
            else:
                first_docid = self.sort_terms[0][0]
                if pivot_doc_id == first_docid:
                    self.current_doc = pivot_doc_id
                    return self.current_doc
                else:
                    #pick all preceding term
                    for i in range(0, pivot_index):
                        change_index = i
                        self.__AdvanceTerm( change_index, pivot_doc_id )
                                                                                    
    def DoQuery(self, query_terms):
        self.__InitQuery(query_terms)
                                                                                        
        while True:
            candidate_docid = self.__Next()
            if candidate_docid == None:
                break
            print "candidate_docid:",candidate_docid
            #insert candidate_docid to heap
            #update threshold
        print "debug_count:",self.debug_count
                                                                                        
if __name__ == "__main__":
    testIndex = {}
    testIndex["t1"] = [ 0, 1, 2, 3, 6 , 2000000000]
    testIndex["t2"] = [ 3, 4, 5, 6, 2000000000 ]
    testIndex["t3"] = [ 2, 5, 2000000000 ]
    testIndex["t4"] = [ 4, 6, 2000000000 ]
    w = WAND(testIndex, 6)
    w.DoQuery(["t1", "t2", "t3", "t4"])

输出结果中会展示next中循环的次数,以及最后被选为candidate的docid。  这里省略了建立堆的过程,使用了一个默认阈值2作为doc的删选条件,候选doc和query doc采用重复词的个数计算UB,这里只是一个算法演示,实际使用的时候需要根据自己的相关性公式进行调整

本文来源:广告系统中weak-and算法原理及编码验证