eda_rowwise.Rmd

# tidyverse中行方向的操作 {#eda-rowwise}

dplyr 1.0 推出之后，数据框**行方向**的操作得到完美解决，因此本章的内容已经过时，大家可以跳出本章，直接阅读第\@ref(tidyverse-colwise) 章。（留着本章，主要是让自己时常回顾下之前的探索。让自己最难忘的，或许就是曾经的痛点吧）


```{r rowwise-1, message = FALSE, warning = FALSE}
library(tidyverse)
```

tidyverse 喜欢数据框，因为一列就是一个向量，一列一列的处理起来很方便。然而我们有时候也要，完成行方向的操作，所以有必要介绍tidyverse中行方向的处理机制。

## 问题

```{r rowwise-2, eval=FALSE}
df <- tibble(x = 1:3, y = 4:6)
df
```
对每行的求和、求均值、最小值或者最大值？


## rowwise函数

dplyr提供了rowwise()函数
```{r rowwise-3, eval=FALSE}
df %>%
  rowwise() %>%
  mutate(i = sum(x, y))
```


```{r rowwise-4, eval=FALSE}
df %>%
  rowwise() %>%
  mutate(i = mean(c(x, y)))
```

```{r rowwise-5, eval=FALSE}
df %>%
  rowwise() %>%
  mutate(
    min = min(x, y),
    max = max(x, y)
  )
```


```{r rowwise-6, eval=FALSE}
df %>%
  rowwise() %>%
  do(i = mean(c(.$x, .$y))) %>%
  unnest(i)
```

## Row-wise Summaries
```{r rowwise-7, eval=FALSE}
df %>% mutate(row_sum = rowSums(.[1:2]))
```


```{r rowwise-8, eval=FALSE}
df %>% mutate(row_mean = rowMeans(.[1:2]))
```

```{r rowwise-9, eval=FALSE}
df %>% mutate(t_sum = rowSums(select_if(., is.numeric)))
```

固然可解决问题， 然而，却不是一个很好的办法，比如除了求和与计算均值，可能还要计算每行的中位数、方差等等， 因为，不是每种计算都对应的row_函数？ 既然是tidyverse ，还是用tidyverse 的方法解决


## purrr::map方案

按照Jenny Bryan的方案

```{r rowwise-10, eval=FALSE}
df %>% mutate(t_sum = pmap_dbl(list(x, y), sum))
```


```{r rowwise-11, eval=FALSE}
df %>%
  mutate(t_sum = pmap_dbl(select_if(., is.numeric), sum))
```

计算均值的时候， 然而报错了
```{r rowwise-12, eval=FALSE}
df %>% mutate(t_sum = pmap_dbl(select_if(., is.numeric), mean))
```

tidyverse 总会想出办法来解决，把`mean()` 变成 `lift_vd(mean)`

```{r rowwise-13, eval=FALSE}
df %>%
  mutate(data = pmap_dbl(select_if(., is.numeric), lift_vd(mean)))
```


同理
```{r rowwise-14, eval=FALSE}
df %>% mutate(t_median = pmap_dbl(select_if(., is.numeric), lift_vd(median)))
```


```{r rowwise-15, eval=FALSE}
df %>% mutate(t_sd = pmap_dbl(select_if(., is.numeric), lift_vd(sd)))
```





## tidy 的方案

我个人推荐的方法(Gather, group, summarize, left_join)

```{r rowwise-16, eval=FALSE}
new_df <- df %>%
  mutate(id = row_number())

s <- new_df %>%
  gather("time", "val", -id) %>%
  group_by(id) %>%
  summarize(
    t_avg = mean(val),
    t_sum = sum(val)
  )

s
```

```{r rowwise-17, eval=FALSE}
new_df %>%
  left_join(s)
```

有点繁琐，但思路清晰

```{r rowwise-18, eval=FALSE}
ss <- new_df %>%
  group_by(id) %>%
  summarise(t_avg = mean(c(x, y)))

ss
```


```{r rowwise-19, eval=FALSE}
new_df %>%
  left_join(ss)
```


之所以有这么多的搞法，是因为没有一个很好的搞法




## 用slide方案

[slide](https://github.com/DavisVaughan/slide)很强大，可以滚动喔


- 如果第一个参数是数据框，`slide`把数据框看作a vector of rows， 然后行方向的滚动，事实上， .x是一个个的小数据框（如下）
- 与`purrr::map`不同，因为map把数据框看作列方向的向量， 然后迭代
- 如果第一个参数是原子型向量的话，还是依次迭代逗号分隔的元素，只不过这里是slide比map更强大的是，还可以是滚动

```{r rowwise-20, eval=FALSE}
library(slider)

df <- tibble(a = 1:3, b = 4:6)

slide(
  select_if(df, is.numeric),
  ~.x,
  .before = 1
)
```



```{r rowwise-21, eval=FALSE}
df %>%
  mutate(
    r_mean = slide_dbl(
      select_if(df, is.numeric),
      ~ mean(unlist(.x)),
      .before = 1
    )
  )
```


##  rowwise() + c_across()

```{r rowwise-22, eval=FALSE}
df <- tibble(id = 1:6, w = 10:15, x = 20:25, y = 30:35, z = 40:45)
df

df %>%
  rowwise(id) %>%
  summarise(mean = mean(c(w, x, y, z)))

df %>%
  rowwise(id) %>%
  mutate(mean = mean(c(w, x, y, z)))

df %>%
  rowwise(id) %>%
  mutate(total = mean(c_across(w:z)))

df %>%
  rowwise(id) %>%
  mutate(mean = mean(c_across(is.numeric)))

# across()
df %>% mutate(mean = rowMeans(across(is.numeric & -id)))
```


## 用lay方案

[lay包](https://github.com/romainfrancois/lay)解决方案
```{r rowwise-23, eval = FALSE}
library(lay)
library(dplyr, warn.conflicts = FALSE)

iris <- as_tibble(iris)

# apply mean to each "row"
iris %>%
  mutate(sepal = lay(across(starts_with("Sepal")), mean))
```


```{r rowwise-24, echo = F}
# remove the objects
# rm(df, new_df, s, ss)
```

```{r rowwise-25, echo = F, message = F, warning = F, results = "hide"}
pacman::p_unload(pacman::p_loaded(), character.only = TRUE)
```