用户标签

代码：用户标签

通过标签将用户和物品联系起来

标签的作用：

1.物品相关：物品的属性（时间，创作者等）

2.用户相关：用户对物品的看法、任务（待读等）

算法：

1.简单算法（推荐用户常用标签下的热门物品）

计算用户对物品的喜好

$p(u,i)=sum_b frac{n_{u,b}}{log(1+n_b^{(u)})}frac{n_{i,b}}{log(1+n_i^{(u)})}$

# coding=gbk
import pandas as pd
import mathdata=pd.read_csv('delicious.dat',sep='	',header=None)class SimpleTagBased:#{用户1:{标签A：5,...}...}user_tag=dict()#用户评价过的物品，在推荐时过滤掉这些物品user_item=dict()#每个标签对应的物品tag_item=dict()tag_count=dict()item_count=dict()item_tag=dict()def __init__(self,data):for user,item,tags in data.itertuples(index=False):if type(tags)==float:continueif user not in self.user_item:self.user_item[user]=list()self.user_item[user].append(item)if item not in self.item_count:self.item_count[item]=0self.item_count[item]+=1if item not in self.item_tag:self.item_tag[item]=dict()tags=tags.split(' ')if user not in self.user_tag:self.user_tag[user]=dict()for tag in tags:tag = tag.lower()if tag not in self.user_tag[user]:self.user_tag[user][tag]=0self.user_tag[user][tag]+=1if tag not in self.tag_item:self.tag_item[tag]=dict()if item not in self.tag_item[tag]:self.tag_item[tag][item]=0self.tag_item[tag][item]+=1if tag not in self.tag_count:self.tag_count[tag]=0self.tag_count[tag]+=1if tag not in self.item_tag[item]:self.item_tag[item][tag]=0self.item_tag[item][tag]+=1def recommend(self,user):viewedItem=self.user_item[user]rank=dict()utags = self.user_tag[user]for tag,weight in utags.items():for item,wt in self.tag_item[tag].items():if item in viewedItem:continueif item not in rank:rank[item]=0#用户user对物品item的喜好程度rank[item]+=weight*1.0/math.log(1+self.tag_count[tag])*wt/math.log(1+self.item_count[item])rank=[a[0] for a in sorted(rank.items(),lambda x,y:cmp(x[1],y[1]),reverse=True)[0:5]];res = []for item in rank:#物品被打的最多的10个标签作为物品的描述res.append([a[0] for a in sorted(self.item_tag[item].items(),lambda x,y:cmp(x[1],y[1]),reverse=True)[0:10]])#用户最常用的10个标签作为用户兴趣描述userdesc = [a[0] for a in sorted(utags.items(),lambda x,y:cmp(x[1],y[1]),reverse=True)[0:10]]return (userdesc,res)
stb = SimpleTagBased(data=data)
userdesc,res = stb.recommend(104)
print userdesc
print res

结果：

--用户常用标签

['software', 'webdesign', 'tools', 'dev', 'howto', 'free', 'freeware', 'opensource', 'reference', 'linux']

--所推荐物品具有的标签

['css', 'webdesign', 'reference', 'design', 'web', 'development', 'html', 'tools', 'webdev', 'programming'],

['webdesign', 'templates', 'design', 'css', 'opensource', 'web', 'free', 'html', 'layout', 'template'],

['fonts', 'typography', 'webdesign', 'design', 'tools', 'css', 'web', 'font', 'type', 'reference'],

['opensource', 'software', 'freeware', 'linux', 'free', 'windows', 'tools', 'reference', 'download', 'alternative'],

['freeware', 'software', 'utilities', 'tools', 'free', 'reference', 'list', 'windows', 'download', 'opensource']

改进：

A.打过标签少的用户（生成相似标签）

标签相似度：（同一物品下的标签相似，两标签同时出现在不同物品下，认为标签相似度高）

使用余弦相似度度量：$(n_{b,1},n_{b,2},n_{b,3},...,n_{b,n})$各量为物品i被打上标签b的次数

# coding=gbk
import pandas as pd
import mathdata=pd.read_csv('delicious.dat',sep='	',header=None)item_tag=dict()
for user,item,tags in data.itertuples(index=False):#如果tags为nan，跳过if type(tags)==float:continueif item not in item_tag:item_tag[item]=dict()tags=tags.split(' ')for tag in tags:tag = tag.lower()if tag not in item_tag[item]:item_tag[item][tag]=0item_tag[item][tag]+=1def recommend(taga,n,item_tag):nb=dict()nab=dict()na = 0l = len(item_tag)i=1for item,tags in item_tag.items():print i*1.0/li+=1if taga not in tags:for tag,v in tags.items():if tag not in nb:nb[tag]=0nb[tag]+=v*velse:av = tags[taga]na +=av*avfor tag,v in tags.items():if tag==taga:continueif tag not in nb:nb[tag]=0nb[tag]+=v*vif tag not in nab:nab[tag]=0nab[tag]+=av*vrank=dict()na = math.sqrt(na)for tag,v in nab.items():rank[tag]=v/na/math.sqrt(nb[tag])res = [a[0] for a in sorted(rank.items(),lambda x,y:cmp(x[1],y[1]),reverse=True)[0:n]]return restags=recommend('webdesign',10,item_tag)
print tags

结果：和webdesign相关的标签

['design', 'css', 'web', 'webdev', 'html', 'web_design', 'inspiration', 'xhtml', 'webdevelopment', 'resources']

2.基于图的算法

解法参看：概率图模型

给用户推荐标签：

对于物品i给用户u推荐标签,标签$b_k$的推荐度为：

$(1-alpha)frac{n_{u,b_k}}{max(n_{u,b_j})}+alphafrac{n_{i,b_k}}{max(n_{i,b_j})}$

# coding=gbk
import pandas as pd
import mathdata=pd.read_csv('delicious.dat',sep='	',header=None)item_tag=dict()
user_tag=dict()
for user,item,tags in data.itertuples(index=False):#如果tags为nan，跳过if type(tags)==float:continueif item not in item_tag:item_tag[item]=dict()if user not in user_tag:user_tag[user]=dict()tags=tags.split(' ')for tag in tags:tag = tag.lower()if tag not in item_tag[item]:item_tag[item][tag]=0item_tag[item][tag]+=1if tag not in user_tag[user]:user_tag[user][tag]=0user_tag[user][tag]+=1def recommend(user,item,user_tag,item_tag,alpha):utgs = user_tag[user]itgs = item_tag[item]udesc = [a[0] for a in sorted(utgs.items(),lambda x,y:cmp(x[1],y[1]),reverse=True)[0:10]]idesc = [a[0] for a in sorted(itgs.items(),lambda x,y:cmp(x[1],y[1]),reverse=True)[0:10]]rank = dict()maxu = max(utgs.values())maxi = max(itgs.values())for tag,v in utgs.items():if tag not in rank:rank[tag] = 0rank[tag]+=(1-alpha)*v/maxufor tag,v in itgs.items():if tag not in rank:rank[tag] = 0rank[tag]+= alpha*v/maxires=[a[0] for a in sorted(rank.items(),lambda x,y:cmp(x[1],y[1]),reverse=True)[0:10]];return udesc,idesc,resudesc,idesc,res=recommend(104,33911,user_tag,item_tag,0.8)
print udesc
print idesc
print res

结果：

--用户常用标签

['software', 'webdesign', 'tools', 'dev', 'howto', 'free', 'freeware', 'opensource', 'reference', 'linux']

--物品常被打标签

['web', 'softwareagents', 'java', 'howto', 'moviles', 'documentation', 'semantica', 'hpi', 'api', 'agents']

--推荐标签

['howto', 'web', 'moviles', 'softwareagents', 'hpi', 'agents', 'api', 'jade', 'agentes', 'java']