社团划分——Fast Unfolding算法

社团划分——Fast Unfolding算法

一、社区划分问题

1、社区以及社区划分

在社交网络中，用户相当于每一个点，用户之间通过互相的关注关系构成了整个网络的结构，在这样的网络中，有的用户之间的连接较为紧密，有的用户之间的连接关系较为稀疏，在这样的的网络中，连接较为紧密的部分可以被看成一个社区，其内部的节点之间有较为紧密的连接，而在两个社区间则相对连接较为稀疏，这便称为社团结构。

(Newman and Gievan 2004) A community is a subgraph containing nodes which are more densely linked to each other than to the rest of the graph or equivalently, a graph has a community structure if the number of links into any subgraph is higher than the number of links between those subgraphs.

如下图：

用红色的点和黑色的点对其进行标注，整个网络被划分成了两个部分，其中，这两个部分的内部连接较为紧密，而这两个社区之间的连接则较为稀疏。如何去划分上述的社区便称为社区划分的问题。

2、社区划分的算法

在社区划分问题中，存在着很多的算法，如由Newman和Gievan提出的GN算法，标签传播算法(Label Propagation Algorithm, LPA)，这些算法都能一定程度的解决社区划分的问题，但是性能则是各不相同。总的来说，在社区划分中，主要分为两大类算法

1. 凝聚方法(agglomerative method)：添加边
2. 分裂方法(divisive method)：移除边

在后续的文章中，我们会继续关注不同的社区划分的算法，在这篇文章中，主要关注Fast Unfolding算法。

3、社区划分的评价标准

为了评价社区划分的优劣，Newman等人提出了模块度的概念，用模块度来衡量社区划分的好坏。简单来讲，就是将连接比较稠密的点划分在一个社区中，这样模块度的值会变大，最终，模块度最大的划分是最优的社区划分。

二、模块度的概念

1、模块度的公式

社区划分的目标是使得划分后的社区内部的连接较为紧密，而在社区之间的连接较为稀疏，通过模块度的可以刻画这样的划分的优劣，模块度越大，则社区划分的效果越好 ，模块度的公式如下所示：

$$Q=\frac{1}{2m}\sum_{i,j}\left [ A_{i,j}-\frac{k_ik_j}{2m} \right ]\delta \left ( c_i,c_j \right )$$

其中，$m = \frac{1}{2}\sum_{i, j}A_{i,j}$表示的是网络中的所有的权重，$A_{i,j}$表示的是节点$i$和节点$j$之间的权重，$k_i=\sum_{j}A_{i,j}$表示的是与顶点$i$连接的边的权重，$c_i$表示的是顶点被分配到的社区，$\delta \left ( c_i,c_j \right )$用于判断顶点$i$与顶点$j$是否被划分在同一个社区中，若是，则返回$1$，否则，返回$0$。

2、模块度公式的简化形式

上述的模块度的计算可以得到以下的简化形式：

$$Q=\sum_{c}\left [ \frac{\sum_{in}}{2m}-\left ( \frac{\sum_ {tot}}{2m} \right )^2 \right ]$$

其中，$\sum_{in}$表示的是社区$c$内部的权重，$\sum_ {tot}$表示的是与社区$c$内部的点连接的边的权重，包括社区内部的边以及社区外部的边。

3、模块度公式的解释

模块度（modularity）指的是网络中连接社区结构内部顶点的边所占的比例，减去在同样的社团结构下任意连接这两个节点的比例的期望值。

三、Fast Unfolding算法

1、Fast Unfolding算法的思路

模块度成为度量社区划分优劣的重要标准，划分后的网络模块度值越大，说明社区划分的效果越好，Fast Unfolding算法便是基于模块度对社区划分的算法，Fast Unfolding算法是一种迭代的算法，主要目标是不断划分社区使得划分后的整个网络的模块度不断增大。

2、Fast Unfolding算法的过程

Fast Unfolding算法主要包括两个阶段，如下图所示：

第一阶段称为Modularity Optimization，主要是将每个节点划分到与其邻接的节点所在的社区中，以使得模块度的值不断变大；第二阶段称为Community Aggregation，主要是将第一步划分出来的社区聚合成为一个点，即根据上一步生成的社区结构重新构造网络。重复以上的过程，直到网络中的结构不再改变为止。

具体的算法过程如下所示：

1. 初始化，将每个点划分在不同的社区中；
2. 对每个节点，将每个点尝试划分到与其邻接的点所在的社区中，计算此时的模块度，判断划分前后的模块度的差值$\Delta Q$是否为正数，若为正数，则接受本次的划分，若不为正数，则放弃本次的划分；
   • 重复以上的过程，直到不能再增大模块度为止；
   • 构造新图，新图中的每个点代表的是步骤3中划出来的每个社区，继续执行步骤2和步骤3，直到社区的结构不再改变为止。

   • 注意：在步骤2中计算节点的顺序对模块度的计算是没有影响的，而是对计算时间有影响。

     四、算法实现

     针对上图表示的网络，最终的结果为：

     

     可以使用下面的程序实现其基本的原理：

     import string
     
     def loadData(filePath):
         f = open(filePath)
         vector_dict = {}
         edge_dict = {}
         for line in f.readlines():
             lines = line.strip().split("\t")
     
             for i in xrange(2):
                 if lines[i] not in vector_dict:
                     #put the vector into the vector_dict
                     vector_dict[lines[i]] = True
                     #put the edges into the edge_dict
                     edge_list = []
                     if len(lines) == 3:
                         edge_list.append(lines[1-i]+":"+lines[2])
                     else:
                         edge_list.append(lines[1-i]+":"+"1")
                     edge_dict[lines[i]] = edge_list
                 else:
                     edge_list = edge_dict[lines[i]]
                     if len(lines) == 3:
                         edge_list.append(lines[1-i]+":"+lines[2])
                     else:
                         edge_list.append(lines[1-i]+":"+"1")
                     edge_dict[lines[i]] = edge_list
     
     return vector_dict, edge_dict
     
     def modularity(vector_dict, edge_dict):
         Q = 0.0
         # m represents the total wight
         m = 0
         for i in edge_dict.keys():
             edge_list = edge_dict[i]
             for j in xrange(len(edge_list)):
                 l = edge_list[j].strip().split(":")
                 m += string.atof(l[1].strip())
     
         # cal community of every vector
         #find member in every community
         community_dict = {}
         for i in vector_dict.keys():
             if vector_dict[i] not in community_dict:
                 community_list = []
             else:
                 community_list = community_dict[vector_dict[i]]
     
             community_list.append(i)
             community_dict[vector_dict[i]] = community_list
     
         #cal inner link num and degree
         innerLink_dict = {}
         for i in community_dict.keys():
             sum_in = 0.0
             sum_tot = 0.0
             #vector num
             vector_list = community_dict[i]
             #print "vector_list : ", vector_list
             #two loop cal inner link
             if len(vector_list) == 1:
                 tmp_list = edge_dict[vector_list[0]]
                 tmp_dict = {}
                 for link_mem in tmp_list:
                     l = link_mem.strip().split(":")
                     tmp_dict[l[0]] = l[1]
                 if vector_list[0] in tmp_dict:
                     sum_in = string.atof(tmp_dict[vector_list[0]])
                 else:
                     sum_in = 0.0
             else:
                 for j in xrange(0,len(vector_list)):
                     link_list = edge_dict[vector_list[j]]
                     tmp_dict = {}
                     for link_mem in link_list:
                         l = link_mem.strip().split(":")
                         #split the vector and weight
                         tmp_dict[l[0]] = l[1]
                     for k in xrange(0, len(vector_list)):
                         if vector_list[k] in tmp_dict:
                             sum_in += string.atof(tmp_dict[vector_list[k]])
     
             #cal degree
             for vec in vector_list:
                 link_list = edge_dict[vec]
                 for i in link_list:
                     l = i.strip().split(":")
                     sum_tot += string.atof(l[1])        
             Q += ((sum_in / m) - (sum_tot/m)*(sum_tot/m))
         return Q
     
     def chage_community(vector_dict, edge_dict, Q):
         vector_tmp_dict = {}
         for key in vector_dict:
             vector_tmp_dict[key] = vector_dict[key]
     
         #for every vector chose it's neighbor
         for key in vector_tmp_dict.keys():
             neighbor_vector_list = edge_dict[key]
             for vec in neighbor_vector_list:
                 ori_com = vector_tmp_dict[key]
                 vec_v = vec.strip().split(":")
     
                 #compare the list_member with ori_com
                 if ori_com != vector_tmp_dict[vec_v[0]]:
                     vector_tmp_dict[key] = vector_tmp_dict[vec_v[0]]
                     Q_new = modularity(vector_tmp_dict, edge_dict)
                     #print Q_new
                     if (Q_new - Q) > 0:
                         Q = Q_new
                     else:
                         vector_tmp_dict[key] = ori_com
         return vector_tmp_dict, Q
     
     def modify_community(vector_dict):
         #modify the community
         community_dict = {}
         community_num = 0
         for community_values in vector_dict.values():
             if community_values not in community_dict:
                 community_dict[community_values] = community_num
                 community_num += 1
         for key in vector_dict.keys():
             vector_dict[key] = community_dict[vector_dict[key]]
         return community_num
     
     def rebuild_graph(vector_dict, edge_dict, community_num):
         vector_new_dict = {}
         edge_new_dict = {}
         # cal the inner connection in every community
         community_dict = {}
         for key in vector_dict.keys():
             if vector_dict[key] not in community_dict:
                 community_list = []
             else:
                 community_list = community_dict[vector_dict[key]]
     
             community_list.append(key)
             community_dict[vector_dict[key]] = community_list
     
         # cal vector_new_dict
         for key in community_dict.keys():
             vector_new_dict[str(key)] = str(key)
     
         # put the community_list into vector_new_dict
     
         #cal inner link num
         innerLink_dict = {}
         for i in community_dict.keys():
             sum_in = 0.0
             #vector num
             vector_list = community_dict[i]
             #two loop cal inner link
             if len(vector_list) == 1:
                 sum_in = 0.0
             else:
                 for j in xrange(0,len(vector_list)):
                     link_list = edge_dict[vector_list[j]]
                     tmp_dict = {}
                     for link_mem in link_list:
                         l = link_mem.strip().split(":")
                         #split the vector and weight
                         tmp_dict[l[0]] = l[1]
                     for k in xrange(0, len(vector_list)):
                         if vector_list[k] in tmp_dict:
                             sum_in += string.atof(tmp_dict[vector_list[k]])
     
             inner_list = []
             inner_list.append(str(i) + ":" + str(sum_in))
             edge_new_dict[str(i)] = inner_list
     
         #cal outer link num
         community_list = community_dict.keys()
         for i in xrange(len(community_list)):
             for j in xrange(len(community_list)):
                 if i != j:
                     sum_outer = 0.0
                     member_list_1 = community_dict[community_list[i]]
                     member_list_2 = community_dict[community_list[j]]
     
                     for i_1 in xrange(len(member_list_1)):
                         tmp_dict = {}
                         tmp_list = edge_dict[member_list_1[i_1]]
     
                         for k in xrange(len(tmp_list)):
                             tmp = tmp_list[k].strip().split(":");
                             tmp_dict[tmp[0]] = tmp[1]
                         for j_1 in xrange(len(member_list_2)):
                             if member_list_2[j_1] in tmp_dict:
                                 sum_outer += string.atof(tmp_dict[member_list_2[j_1]])
     
                     if sum_outer != 0:
                         inner_list = edge_new_dict[str(community_list[i])]
                         inner_list.append(str(j) + ":" + str(sum_outer))
                         edge_new_dict[str(community_list[i])] = inner_list
         return vector_new_dict, edge_new_dict, community_dict
     
     def fast_unfolding(vector_dict, edge_dict):
         #1. initilization:put every vector into different communities
         #   the easiest way:use the vector num as the community num
         for i in vector_dict.keys():
             vector_dict[i] = i
     
         #print "vector_dict : ", vector_dict
         #print "edge_dict : ", edge_dict
     
         Q = modularity(vector_dict, edge_dict)  
     
         #2. for every vector, chose the community
         Q_new = 0.0
         while (Q_new != Q):
             Q_new = Q
             vector_dict, Q = chage_community(vector_dict, edge_dict, Q)
         community_num = modify_community(vector_dict)
         print "Q = ", Q
         print "vector_dict.key : ", vector_dict.keys()
         print "vector_dict.value : ", vector_dict.values()
         Q_best = Q
         while (True):
             #3. rebulid new graph, re_run the second step
             print "edge_dict : ",edge_dict
             print "vector_dict : ",vector_dict
             print "\n rebuild"
             vector_dict, edge_new_dict, community_dict = rebuild_graph(vector_dict, edge_dict, community_num)
             #print vector_dict
             print "community_dict : ", community_dict
     
             Q_new = 0.0
             while (Q_new != Q):
                 Q_new = Q
                 vector_dict, Q = chage_community(vector_dict, edge_new_dict, Q)
             community_num = modify_community(vector_dict)
             print "Q = ", Q
             if (Q_best == Q):
                 break
             Q_best = Q
             vector_result = {}
             for key in community_dict.keys():
                 value_of_vector = community_dict[key]
                 for i in xrange(len(value_of_vector)):
                     vector_result[value_of_vector[i]] = str(vector_dict[str(key)])
             for key in vector_result.keys():
                 vector_dict[key] = vector_result[key]
             print "vector_dict.key : ", vector_dict.keys()
             print "vector_dict.value : ", vector_dict.values()
     
         #get the final result
         vector_result = {}
         for key in community_dict.keys():
             value_of_vector = community_dict[key]
             for i in xrange(len(value_of_vector)):
                 vector_result[value_of_vector[i]] = str(vector_dict[str(key)])
         for key in vector_result.keys():
             vector_dict[key] = vector_result[key]
         print "Q_best : ", Q_best
         print "vector_result.key : ", vector_dict.keys()
         print "vector_result.value : ", vector_dict.values()
     
     if __name__ == "__main__":
         vector_dict, edge_dict=loadData("./cd_data.txt")
     
         fast_unfolding(vector_dict, edge_dict)
     

     ---

     参考文献

     1. Vincent D Blondel, Jean-Loup Guillaume, Renaud Lambiotte, Etienne Lefebvre, Fast unfolding of communities in large networks, in Journal of Statistical Mechanics: Theory and Experiment 2008 (10), P1000
     2. 社区发现算法FastUnfolding的GraphX实现 http://www.tuicool.com/articles/Jrieue