通过填充转发/ LOCF在SQL中的一系列连续行上估算一列?

Dav*_*idJ 5 sql time-series data-warehouse data-cleaning

在格式化数据以进行时间序列分析时,通常需要通过随时间推移填充前馈值(也称为Last-Observation-Carried-Forward / LOCF)来估算缺失值。

尽管数据分析环境通常提供该功能(例如Pandas fillna()),但对于较大的数据集,在SQL中进行计算(例如,利用数据并行数据仓库设备)可能会更加高效。

例如,考虑:

    | UNIT | TIME | VALUE |
    |------|------|-------|
    | 1    | 1    | NULL  |
    | 1    | 2    | .5    |
    | 1    | 3    | NULL  |
    | 1    | 4    | NULL  |
    | 1    | 5    | .2    |
    | 1    | 6    | NULL  |
    | 2    | 1    | .6    |
    | 2    | 2    | NULL  |
Run Code Online (Sandbox Code Playgroud)

在TIME之前填充VALUE列后(对于每个UNIT独立),结果如下:

    | UNIT | TIME | VALUE |
    |------|------|-------|
    | 1    | 1    | NULL  |
    | 1    | 2    | .5    |
    | 1    | 3    | .5    |
    | 1    | 4    | .5    |
    | 1    | 5    | .2    |
    | 1    | 6    | .2    |
    | 2    | 1    | .6    |
    | 2    | 2    | .6    |
Run Code Online (Sandbox Code Playgroud)

(请注意,由于没有先验值,因此无法估算UNIT 1的初始NULL)

时间也可以是时间戳记或日期时间类型列。

Vin*_*ynd 9

对于某些数据库,例如 Postgres,您可以定义自己的聚合函数。LOCF 只是一个正在运行的 COALESCE。

CREATE OR REPLACE FUNCTION locf_state( FLOAT, FLOAT )
RETURNS FLOAT
LANGUAGE SQL
AS $f$
  SELECT COALESCE($2,$1)
$f$;

CREATE AGGREGATE locf(FLOAT) (
  SFUNC = locf_state,
  STYPE = FLOAT
);
Run Code Online (Sandbox Code Playgroud)

这样查询就更具可读性:

SELECT unit, time, 
       locf(value) OVER( PARTITION BY unit ORDER BY time )
FROM   mytable;
Run Code Online (Sandbox Code Playgroud)

SQLFiddle: http: //sqlfiddle.com/#!15/2c73b/ 1/0


Dav*_*idJ 2

如果使用 PostgreSQL 风格的 SQL 方言(例如 Netezza PureData)作为日期时间索引(假设过去的数据),以下查询结构将实现前向填充。它也适用于多列索引/键。

给定以下参数:

  • <key_cols>- 唯一标识每个时间序列样本的列列表(例如UNIT, TIME
  • <impute_col>- 需要估算值的列(例如VALUE
  • <impute_over_range_col>- 时间序列的连续范围列(例如TIME

并推导:

  • <keys_no_range>- 关键列,除了<impute_over_range_col>

SELECT DISTINCT T1.<key_cols>, 
                COALESCE(T1.<impute_col>, T2.<impute_col>) AS <impute_col>
FROM table T1
LEFT OUTER JOIN (SELECT T1.<key_cols>,
                     T1.<impute_col>,
                     LEAD(T1.<impute_over_range_col>,1) 
                         OVER (PARTITION BY T1.<keys_no_range> 
                               ORDER BY T1.<key_cols>)
                         AS NEXT_RANGE
                     FROM table T1
                     WHERE T1.<impute_col> IS NOT NULL
                     ORDER BY T1.<key_cols>
                ) T2
              ON (T1.<impute_over_range_col> BETWEEN T2.<impute_over_range_col> 
                                             AND COALESCE(NEXT_RANGE, CURRENT_DATE)) 
              AND T1.<keys_no_range>[0] = T2.<keys_no_range>[0]
              AND T1.<keys_no_range>[1] = T2.<keys_no_range>[1]
              -- ... for each col in <keys_no_range>
Run Code Online (Sandbox Code Playgroud)

具体来说,对于问题中的例子:

SELECT DISTINCT T1.UNIT, T1.TIME, 
                COALESCE(T1.VALUE, T2.VALUE) AS VALUE
FROM table T1
LEFT OUTER JOIN (SELECT T1.UNIT, T1.TIME,
                     T1.VALUE,
                     LEAD(T1.TIME,1) 
                         OVER (PARTITION BY T1.UNIT 
                               ORDER BY T1.UNIT, T1.TIME)
                         AS NEXT_RANGE
                     FROM table T1
                     WHERE T1.VALUE IS NOT NULL
                     ORDER BY T1.UNIT, T1.TIME
                ) T2
              ON (T1.TIME BETWEEN T2.TIME
                           AND COALESCE(NEXT_RANGE, CURRENT_DATE)) 
              AND T1.UNIT = T2.UNIT
Run Code Online (Sandbox Code Playgroud)

这是上述查询的 SQLFiddle:http://sqlfiddle.com/#!15 /d589b/1