R for Data Science总结之——modelr(2)
R for Data Science总结之——modelr(2)
本章针对真实数据集进行建模实践:
library(tidyverse)
library(modelr)
options(na.action = na.warn)
library(nycflights13)
library(lubridate)
为什么低质量的钻石更加昂贵?
首先查看diamonds数据集:
ggplot(diamonds, aes(cut, price)) + geom_boxplot()
ggplot(diamonds, aes(color, price)) + geom_boxplot()
ggplot(diamonds, aes(clarity, price)) + geom_boxplot()
再查看连续变量克拉与价格的关系:
ggplot(diamonds, aes(carat, price)) +
geom_hex(bins = 50)
再针对小于2.5克拉的钻石,抛弃极值点:
diamonds2 <- diamonds %>%
filter(carat <= 2.5) %>%
mutate(lprice = log2(price), lcarat = log2(carat))
ggplot(diamonds2, aes(lcarat, lprice)) +
geom_hex(bins = 50)
对对数化数据进行建模:
mod_diamond <- lm(lprice ~ lcarat, data = diamonds2)
grid <- diamonds2 %>%
data_grid(carat = seq_range(carat, 20)) %>%
mutate(lcarat = log2(carat)) %>%
add_predictions(mod_diamond, "lprice") %>%
mutate(price = 2 ^ lprice)
ggplot(diamonds2, aes(carat, price)) +
geom_hex(bins = 50) +
geom_line(data = grid, colour = "red", size = 1)
再查看残差:
diamonds2 <- diamonds2 %>%
add_residuals(mod_diamond, "lresid")
ggplot(diamonds2, aes(lcarat, lresid)) +
geom_hex(bins = 50)
再查看之前的几个变量与残差的关系:
ggplot(diamonds2, aes(cut, lresid)) + geom_boxplot()
ggplot(diamonds2, aes(color, lresid)) + geom_boxplot()
ggplot(diamonds2, aes(clarity, lresid)) + geom_boxplot()
可以看出这三个变量都与残差有关,因此建立复杂模型:
mod_diamond2 <- lm(lprice ~ lcarat + color + cut + clarity, data = diamonds2)
然后使用data_grid()的.model参数修改过程,这里由于R版本问题有时会报错,若报错请查看博客底部GITHUB的详细代码:
grid <- diamonds2 %>%
data_grid(cut, .model = mod_diamond2) %>%
add_predictions(mod_diamond2)
grid
#> # A tibble: 5 x 5
#> cut lcarat color clarity pred
#> <ord> <dbl> <chr> <chr> <dbl>
#> 1 Fair -0.515 G VS2 11.2
#> 2 Good -0.515 G VS2 11.3
#> 3 Very Good -0.515 G VS2 11.4
#> 4 Premium -0.515 G VS2 11.4
#> 5 Ideal -0.515 G VS2 11.4
ggplot(grid, aes(cut, pred)) +
geom_point()
再查看残差:
diamonds2 <- diamonds2 %>%
add_residuals(mod_diamond2, "lresid2")
ggplot(diamonds2, aes(lcarat, lresid2)) +
geom_hex(bins = 50)
我们再单独查看残差极大的值:
diamonds2 %>%
filter(abs(lresid2) > 1) %>%
add_predictions(mod_diamond2) %>%
mutate(pred = round(2 ^ pred)) %>%
select(price, pred, carat:table, x:z) %>%
arrange(price)
#> # A tibble: 16 x 11
#> price pred carat cut color clarity depth table x y z
#> <int> <dbl> <dbl> <ord> <ord> <ord> <dbl> <dbl> <dbl> <dbl> <dbl>
#> 1 1013 264 0.25 Fair F SI2 54.4 64 4.3 4.23 2.32
#> 2 1186 284 0.25 Premium G SI2 59 60 5.33 5.28 3.12
#> 3 1186 284 0.25 Premium G SI2 58.8 60 5.33 5.28 3.12
#> 4 1262 2644 1.03 Fair E I1 78.2 54 5.72 5.59 4.42
#> 5 1415 639 0.35 Fair G VS2 65.9 54 5.57 5.53 3.66
#> 6 1415 639 0.35 Fair G VS2 65.9 54 5.57 5.53 3.66
#> # ... with 10 more rows
什么影响每天的航班数量?
daily <- flights %>%
mutate(date = make_date(year, month, day)) %>%
group_by(date) %>%
summarise(n = n())
daily
#> # A tibble: 365 x 2
#> date n
#> <date> <int>
#> 1 2013-01-01 842
#> 2 2013-01-02 943
#> 3 2013-01-03 914
#> 4 2013-01-04 915
#> 5 2013-01-05 720
#> 6 2013-01-06 832
#> # ... with 359 more rows
ggplot(daily, aes(date, n)) +
geom_line()
我们再看每一周的航班数量分布:
daily <- daily %>%
mutate(wday = wday(date, label = TRUE))
ggplot(daily, aes(wday, n)) +
geom_boxplot()
可以看出周六航班极为稀少,下一步建模:
mod <- lm(n ~ wday, data = daily)
grid <- daily %>%
data_grid(wday) %>%
add_predictions(mod, "n")
ggplot(daily, aes(wday, n)) +
geom_boxplot() +
geom_point(data = grid, colour = "red", size = 4)
计算残差:
daily <- daily %>%
add_residuals(mod)
daily %>%
ggplot(aes(date, resid)) +
geom_ref_line(h = 0) +
geom_line()
可以看出有一些拟合极度糟糕的情况,接下来按每周进行分类:
ggplot(daily, aes(date, resid, colour = wday)) +
geom_ref_line(h = 0) +
geom_line()
可以看出周六的预测完全失败了,我们挑出极值点:
daily %>%
filter(resid < -100)
#> # A tibble: 11 x 4
#> date n wday resid
#> <date> <int> <ord> <dbl>
#> 1 2013-01-01 842 Tue -109.
#> 2 2013-01-20 786 Sun -105.
#> 3 2013-05-26 729 Sun -162.
#> 4 2013-07-04 737 Thu -229.
#> 5 2013-07-05 822 Fri -145.
#> 6 2013-09-01 718 Sun -173.
#> # ... with 5 more rows
可以看出这里包括一些美国节日,这是导致航班数量与模型不吻合的原因。
daily %>%
ggplot(aes(date, resid)) +
geom_ref_line(h = 0) +
geom_line(colour = "grey50") +
geom_smooth(se = FALSE, span = 0.20)
#> `geom_smooth()` using method = 'loess' and formula 'y ~ x'
我们再观察一下季节性与星期间的关系:
daily %>%
filter(wday == "Sat") %>%
ggplot(aes(date, n)) +
geom_point() +
geom_line() +
scale_x_date(NULL, date_breaks = "1 month", date_labels = "%b")
这里我们创建term()函数来包含三个学期:
term <- function(date) {
cut(date,
breaks = ymd(20130101, 20130605, 20130825, 20140101),
labels = c("spring", "summer", "fall")
)
}
daily <- daily %>%
mutate(term = term(date))
daily %>%
filter(wday == "Sat") %>%
ggplot(aes(date, n, colour = term)) +
geom_point(alpha = 1/3) +
geom_line() +
scale_x_date(NULL, date_breaks = "1 month", date_labels = "%b")
再看下星期间的差别:
daily %>%
ggplot(aes(wday, n, colour = term)) +
geom_boxplot()
这里看出学期中的区别时模型拟合不准确的一大因素,因此创建模型:
mod1 <- lm(n ~ wday, data = daily)
mod2 <- lm(n ~ wday * term, data = daily)
daily %>%
gather_residuals(without_term = mod1, with_term = mod2) %>%
ggplot(aes(date, resid, colour = model)) +
geom_line(alpha = 0.75)
可以看出残差明显缩小了,再观察拟合结果:
grid <- daily %>%
data_grid(wday, term) %>%
add_predictions(mod2, "n")
ggplot(daily, aes(wday, n)) +
geom_boxplot() +
geom_point(data = grid, colour = "red") +
facet_wrap(~ term)
我们可以发现拟合结果较好,但仍有极值点的存在,因此可使用RGM也就是MASS::rlm()消除极值点影响:
mod3 <- MASS::rlm(n ~ wday * term, data = daily)
daily %>%
add_residuals(mod3, "resid") %>%
ggplot(aes(date, resid)) +
geom_hline(yintercept = 0, size = 2, colour = "white") +
geom_line()
可以看出除特殊日期外的残差明显减小了。
这里需要注意,如果项进行多次可视化,则需要计算新的变量,推荐使用函数进行计算:
compute_vars <- function(data) {
data %>%
mutate(
term = term(date),
wday = wday(date, label = TRUE)
)
}
或者直接在模型中进行计算:
wday2 <- function(x) wday(x, label = TRUE)
mod3 <- lm(n ~ wday2(date) * term(date), data = daily)
也可以用多项式拟合:
library(splines)
mod <- MASS::rlm(n ~ wday * ns(date, 5), data = daily)
daily %>%
data_grid(wday, date = seq_range(date, n = 13)) %>%
add_predictions(mod) %>%
ggplot(aes(date, pred, colour = wday)) +
geom_line() +
geom_point()
这里也推荐学习caret包进行高级建模The caret Package
全文代码已上传GITHUB点此进入