BigQuery中交叉连接后的行聚合

Question

假设您在BigQuery中有以下表格:

A = user1 | 0 0 |
    user2 | 0 3 |
    user3 | 4 0 |

交叉加入后,你有

dist = |user1  user2  0 0 , 0 3 |  #comma is just showing user val seperation
       |user1  user3  0 0 , 4 0 |
       |user2  user3  0 3 , 4 0 |

如何在BigQuery中执行行聚合以计算跨行的成对聚合.作为典型用例,您可以计算两个用户之间的欧氏距离.我想在两个用户之间计算以下指标:

sum(min(user1_row[i], user2_row[i]) / abs(user1_row[i] - user2_row[i]))

为每对用户总结了所有i.

例如,在Python中,您只需:

for i in np.arange(row_length/2)]):
dist.append([user1, user2, np.sum(min(r1[i], r2[i]) / abs(r1[i] - r2[i]))])

Answer 1

从丑陋的方式开始:你可以将数学变成查询.也就是说,转换 for i in ... sum(min(...)/abs(...))为在每个字段上运行的SQL.请注意,MIN并且SUM是您不想使用的聚合函数.而是+用于SUM和IF(a < b, a, b)for MIN.ABS(a, b)看起来像IF(a < b, b-a, a-b).如果你只是计算欧几里德距离,你可以做到

SELECT left.user, right.user, 
  SQRT((left.x-right.x)*(left.x-right.x) 
     + (left.y-right.y)*(left.y-right.y) 
     + (left.z-right.z)*(left.z-right.z)) as dist 
FROM (
    SELECT * 
    FROM dataset.table1 AS left 
    CROSS JOIN dataset.table1 AS right)

更好的方法是用户定义函数,并将向量创建为重复值.然后,您可以编写一个DISTANCE()函数,对交叉连接的左侧和右侧的两个数组执行计算.如果您不在UDF测试计划中并想加入,请联系google云支持.

最后,如果您将架构更改{user:string, field1:float, field2:float, field3:float,...}为{user:string, fields:[field:float]}

然后,您可以使用位置展平该字段并对其进行交叉连接.如:

SELECT
  user,
  field,
  index,
FROM (FLATTEN((
  SELECT  
    user,
    fields.field as field,
    POSITION(fields.field) as index,
  from [dataset1.table1]
), fields))

如果将其另存为视图,请将其命名为"dataset1.flat_view"

然后你可以加入:

SELECT left.user as user1, right.user as user2,
       left.field as l, right.field as r,
FROM dataset1.flat_view left
JOIN dataset1.flat_view right 
ON left.index = right.index
WHERE left.user != right.user

这将为每对用户和每个字段匹配字段分别提供一行.您可以将其保存为视图"dataset1.joined_view".

最后,您可以进行聚合:

既然你想要这个:

sum(min(user1_row[i], user2_row[i]) / abs(user1_row[i] - user2_row[i]))

它看起来像:

SELECT user1, user2, 
    SUM((if (l < r, l, r)) / (if (l > r, l-r, r-l))
FROM [dataset1.joined_view] 
GROUP EACH BY user1, user2