逻辑斯蒂回归用于二分类0/1 因变量因变量 y只能取 0 或 1事件发生概率范围 (0,1)# 加载内置数据集mtcars二分类变量vs0V型1直列 data(mtcars) # 查看数据 head(mtcars) # 目标用mpg、wt、hp预测 vs0/1二分类 # glm() 通用线性模型family binomial 就是逻辑斯蒂回归 logit_model - glm( formula vs ~ mpg wt hp, # 因变量~自变量 data mtcars, family binomial(link logit) ) # 输出模型详细结果 summary(logit_model)Call:glm(formula vs ~ mpg wt hp, family binomial(link logit), data mtcars)Deviance Residuals: # 偏差残差越小拟合越好Min 1Q Median 3Q Max-2.0157 -0.4043 -0.0158 0.4775 1.8834Coefficients: # 核心系数表Estimate Std. Error z value Pr(|z|)(Intercept) 18.57445 7.38663 2.515 0.01192 *mpg -0.11965 0.14794 -0.809 0.41879wt -4.24346 1.86899 -2.270 0.02317 *hp -0.03237 0.01821 -1.778 0.07539 .---Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1Estimate回归系数(beta正数自变量越大(y1)概率越大负数自变量越大(y1)概率越小Std.Error系数标准误z valueZ 检验统计量Pr(|z|)P 值(P0.05)自变量显著影响分类结果# 提取优势比 exp(系数) exp(coef(logit_model)) # 置信区间 exp(confint(logit_model))Waiting for profiling to be done...2.5 % 97.5 %(Intercept) 1.720775e-22 1.062851e09mpg 7.432421e-01 5.813725e00wt 2.934997e-01 2.461978e05hp 8.035275e-01 9.713645e-01共有25个警告 (用warnings()来显示)(OR1)自变量增大发生事件概率上升(OR1)自变量增大发生事件概率下降预测与评价# 1. 训练集内预测概率 pre_prob - predict(logit_model, type response) # 2. 转为0/1分类结果阈值0.5 pre_class - ifelse(pre_prob 0.5, 1, 0) head(pre_class) # 真实值 vs 预测值 table(真实值mtcars$vs, 预测值pre_class) # 计算准确率 acc - mean(pre_class mtcars$vs) acc预测值真实值 0 10 16 21 0 14 # 计算准确率[1] 0.9375可视化# 1. 生成自变量序列连续取值用于画平滑曲线 wt_seq - seq(min(mtcars$wt), max(mtcars$wt), length.out 100) # 2. 构造预测数据集其余自变量取均值控制变量 new_data - data.frame( wt wt_seq, mpg mean(mtcars$mpg), hp mean(mtcars$hp) ) # 3. 预测概率 prob_pred - predict(logit_model, newdata new_data, type response) # 4. 绘图原始散点 拟合S曲线 plot( x mtcars$wt, y mtcars$vs, main 逻辑斯蒂回归 S型概率曲线(车重vs分类), xlab 车重 wt, ylab 分类 vs (0/1), pch 16, col steelblue ) # 绘制拟合曲线 lines(wt_seq, prob_pred, lwd 2, col red) # 添加0.5阈值线 abline(h 0.5, lty 2, col gray) legend(topright, legend c(原始数据, 拟合概率曲线, 0.5阈值), col c(steelblue, red, gray), pch c(16, NA, NA), lty c(NA, 1, 2))