forked from TsingH-googl/RNR_framework_for_Recommendation
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathrandomTest.m
143 lines (130 loc) · 6.34 KB
/
randomTest.m
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
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
clear all;
t1=clock();
%% 读入user-item二分图数据
% 非时序网络：
% D:\链路预测相关课题\基于用户-物品的推荐\无时间上下文信息数据集\unicodelang\out.unicodelang， 614*254，需要转置
% D:\链路预测相关课题\基于用户-物品的推荐\无时间上下文信息数据集\moreno_crime\out.moreno_crime_crime， 829*551，需要转置
% D:\链路预测相关课题\基于用户-物品的推荐\无时间上下文信息数据集\brunson_revolution\out.brunson_revolution_revolution， 5*136， 需要转置
% 时序网络：
% D:\链路预测相关课题\基于用户-物品的推荐\时间上下文信息数据集\MovieLens\ml-100k\u.data， 943*1682, 15s
% D:\链路预测相关课题\基于用户-物品的推荐\时间上下文信息数据集\MovieLens\ml-1m\data.txt， 6000*4000， 4.5mins（改进前），3mins（改进后）
% D:\链路预测相关课题\基于用户-物品的推荐\时间上下文信息数据集\edit-frwikibooks\out.edit-frwikibooks，需要转置， 28113*2884
% D:\链路预测相关课题\基于用户-物品的推荐\时间上下文信息数据集\escorts\out.escorts， 10000*6000，2.4mins
% D:\链路预测相关课题\基于用户-物品的推荐\时间上下文信息数据集\lastfm-2k\user_taggedartists-timestamps.dat, 2000*19000
X = load('D:\链路预测相关课题\基于用户-物品的推荐\时间上下文信息数据集\MovieLens\ml-100k\u.data');
% temp=X(:,1);X(:,1)=X(:,2);X(:,2)=temp;%某些konect中的bipartite的user与item列是调换了的
data = X(:,[1 2 3 4]);
net = spconvert(data(:,[1 2 3])); % net包含rating信息
%% PNR模型算法参数
pH = 0.1; % 测试集占比
winSize = 7; % PNR滤波窗口大小
bin = 80; % 计算PNR的bin数目
topL = 10; % topL推荐：对“每个用户”都推荐topL个item，有趣的现象：给他们推荐topL"大于"low_degree则会有明显效果
mu = 0; % 正态分布均值
sigma = 1; % 正态分布方差
train_start_ratio = 0.0;
train_end_ratio = 0.6;
probe_start_ratio = 0.6;
probe_end_ratio = 0.7;
% low_degree = 50; % 与topL联系起来，low_degree增大PNR变优会影响整体性能
%% CF参数
K = 25; % 计算user-item similarity的参数，对结果的影响也很大，K增大会使得PNR结果变优！！K太小则呈山峰爬升状而无效果
% wikibooks取35
%% basicMF参数
basicMF_lambda = 10;
basicMF_feat_num = 10;
basicMF_maxiter = 20;
%% FISM参数
FISM_n_lfactors = 96;% ml-100k为96
FISM_alpha = 0.1; % ml-100k为0.4
%% IBPR参数
IBPR_maxIter = 50;% 迭代次数，一般是25的整数倍
IBPR_d = 30; % U、V特征向量的维度数
%% MatricCompletion
MC_mu = 0.01;% mu = [0.001,0.01]
MC_rho = 5;% rho = [1,5]
MC_maxiter = 20;
%% MFSGD参数
iteraTime = 10; % 计算MF-SGD的迭代次数
factor = 30; % 计算MF-SGD的矩阵分解因子
num_feature = 100; % RankSGD特征数目
%% PureSVD的参数
svd_k = 500; % 计算PureSVD的参数,ml100k=300、200则有效
svd_maxit = 20;
%% NMF参数
nmf_k = 30;
nmf_maxit = 50;
%% PMF参数
theta = 5; % 不确定是啥
lambda_u = 0.01;
lambda_v = 0.001;
l= 0.005; % 不确定是不是learningRate
num_iters = 10;
%% 时序网络划分数据集,0-1图
% [train, probe] = DivideNet(net, pH);
[train, probe] = timeDivideSeg(data, train_start_ratio, train_end_ratio, probe_start_ratio, probe_end_ratio);
% deg_index = find(sum(train, 2) > 0); % 去掉user。在cosine_user下去掉小度用户后，效果好很多!!
% train = train(deg_index,:);
% probe = probe(deg_index,:);
% net = net(deg_index,:);
% deg_index = find(sum(train, 1) > 0); % 去掉item
% train = train(:, deg_index);
% probe = probe(:, deg_index);
% net = net(:, deg_index);
train_rating = train .* net;
probe_rating = probe .* net;
%% 计算user-item的score，接着计算PNR
% sim = cosine_item_similarity(train);
% score = itemCF( train, sim, K);
score = MatrixCompletion( train ,MC_mu, MC_rho, MC_maxiter);
% score = FISM( train, FISM_n_lfactors, FISM_alpha);
% score = (IBPR( train_rating, IBPR_maxIter, IBPR_d)); % recall与precision好像出问题，要注意使用！！
% score = basicMF( train_rating, basicMF_lambda , basicMF_feat_num, basicMF_maxiter);
% score = MFSGD(train, iteraTime, factor);
% [U,V] = PMF(R,theta,lambda_u,lambda_v,l,num_iters);
% score = U * V';
% score = NMF( train, nmf_k, nmf_maxit);
% score = abs( PureSVD( train , svd_k, svd_maxit) ); % 加绝对值区别很大！！一定要加！！
result = hybridDistribution (score, train, bin, [0]); % 把score矩阵所有的0去掉（有时候不去掉0也会有更好的效果）
alpha = nnz(probe)/ ( size(train, 1) * size(train, 2) - nnz(train) ); % 常数，使得PNR纵坐标的值没那么大
pnr = PNR(result(:,5), result(:,6), result(:,1) , alpha);
%% 对PNR进行滤波
y = pnr(:,2)';
pnrGau = [pnr(:,1), smoothts(y,'g',winSize)']; % 高斯滤波，线性。
pnrMed = [pnr(:,1), medfilt1(y, winSize)']; % 中值滤波，非线性。
pnrMean = [pnr(:,1), meanFilter(y, winSize)']; % 均值滤波，线性。
pnrExp = [pnr(:,1), smoothts(y,'e',winSize)']; % 指数法滤波。
%% 修改分数
% scorePNR = readjustScore (score, pnr);
% scoreGau = readjustScore (score, pnrGau);
% scoreMed = readjustScore (score, pnrMed);
% scoreMean = readjustScore(score, pnrMean);
% scoreExp = readjustScore (score, pnrExp);
scorePNR = readjustScoreUnique (score, pnr, mu, sigma);% wikibooks数据集bin1/2/3变为0会有超级结果
scoreGau = readjustScoreUnique (score, pnrGau, mu, sigma);
scoreMed = readjustScoreUnique (score, pnrMed, mu, sigma);
scoreMean = readjustScoreUnique(score, pnrMean, mu, sigma);
scoreExp = readjustScoreUnique (score, pnrExp, mu, sigma);
%% “分类准确度”评估指标：准确率（precision）、召回率(recall)（F-Measure，ep(L)，er(L)评估指标都是基于这两者）
[p, r, hr, arhr] = evaluator( train ,probe, score, topL);
[pPNR, rPNR, hrPNR, arhrPNR] = evaluator( train ,probe, scorePNR, topL);
[pGau, rGau, hrGau, arhrGau] = evaluator( train ,probe, scoreGau, topL);
[pMed, rMed, hrMed, arhrMed] = evaluator( train ,probe, scoreMed, topL);
[pMean, rMean, hrMean, arhrMean] = evaluator( train ,probe, scoreMean,topL);
[pExp, rExp, hrExp, arhrExp] = evaluator( train ,probe, scoreExp, topL);
plotPNR(pnr(:,1), [pnr(:,2),pnrGau(:,2),pnrMed(:,2),pnrMean(:,2),pnrExp(:,2)]);
%% “分类准确度”评估指标：AUC (不知道为啥这里有错误)
% aucPNR = CalcAUC(train, probe, scorePNR);
% aucGau = CalcAUC(train, probe, scoreGau);
% aucMed = CalcAUC(train, probe, scoreMed);
% aucMean = CalcAUC(train, probe, scoreMean);
% aucExp = CalcAUC(train, probe, scoreExp);
% auc = CalcAUC(train, probe, score);
%% “排序准确度”：平均排序分（average ranking score），小
RS = rankingScore( train, probe, score);
RS_PNR = rankingScore( train, probe, scorePNR);
RS_Mean = rankingScore( train, probe, scoreMean);
RS_Exp = rankingScore( train, probe, scoreExp);
RS_Gau = rankingScore( train, probe, scoreGau);
RS_Med = rankingScore( train, probe, scoreMed);
t2=clock();