BigQueryで統計量を出す時に使うクエリメモ

これは何か

window関数を駆使するとUDFを使わなくとも簡潔に統計量が出せる。がよく書き方を忘れるのでメモ

基本

-- 平均
AVG(expr)

-- 中央値0.1%誤差
NTH(501, QUANTILES(expr, 1001))

-- 最頻値
TOP(expr, 1), COUNT(*) as count

-- 分散
VARIANCE(expr) 
-- 母集団の分散 (分母が n-1 の方)
VARIANCE_POP(expr)

-- 標準偏差
STDDEV(expr)
-- 母集団の標準偏差 (分母が n-1 の方)
STDDEV_POP(expr)

-- 変動係数
STDDEV(expr)/AVG(expr)

-- ピアソン相関係数
CORR(expr1, expr2)

情報量

Public DatasetsのUSA Namesから、ジェンダー別に名前毎の情報量

select
	gender,
	name,
	number,
	-log2(p_name) as bits
from (
	select
		gender,
		name,
		number,
		ratio_to_report(number) over (partition by gender) as p_name,
	from [bigquery-public-data:usa_names.usa_1910_2013]
	where state = 'CA' 
)

エントロピー

Public DatasetsのUSA Namesから、ジェンダー毎に名前のエントロピー

select
	state,
	gender,
	sum(p_name * -log2(p_name)) as entropy
from (
	select
		state,
		gender,
		name,
		number,
		ratio_to_report(number) over (partition by gender) as p_name,
	from [bigquery-public-data:usa_names.usa_1910_2013]
)
group by state, gender
order by entropy

KL-Divergence

KLDivergence = \sum_iP(i)log_2\frac{P(i)}{Q(i)}

KaggleのClick-Through Rate Predictionのデータを例に、広告枠毎にhourlyのインプレッション数とクリック数の確率分布の距離を取る

select
    site_id,
    sum(p_imp * log2(p_imp/(p_click + 0.0001))) as KLD_imp_and_click
from (
    select 
        site_id,
        hour,
        imp,
        click,
        ratio_to_report(imp) over (partition by site_id) as p_imp,
        ratio_to_report(click) over (partition by site_id) as p_click
    from (
        select 
            site_id,
            hour,
            count(*) as imp,
            sum(click) as click,
        from [hagino3000-bq-sandbox:ctr_competition.train]
        group by site_id, hour
        order by site_id, hour
    )
)
group by site_id
order by KLD_imp_and_click

時系列データの移動平均と差分系列

同じくKaggleのClick-Through Rate Predictionで広告枠ごとの3時間移動平均。欠損値を補完したい場合は軸を別途作ってJoinする事。

select
  site_id,
  hour,
  avg(imp) over (partition by site_id
                 order by hour
                 ROWS BETWEEN 3 preceding AND CURRENT ROW) as rolling_mean
from (
  select 
      site_id,
      hour,
      count(*) as imp
  from [hagino3000-bq-sandbox:ctr_competition.train]
      group by site_id, hour
)
order by site_id, hour

差分系列

select
    site_id,
    hour,
    imp,
    imp - ifnull(previous_imp, imp) as residual
from (
    select
        site_id,
        hour,
        imp,
        lag(imp, 1) over (partition by site_id order by hour) as previous_imp
    from (
        select 
            site_id,
            hour,
            count(*) as imp
        from [hagino3000-bq-sandbox:ctr_competition.train]
        group by site_id, hour
    )
)
order by site_id, hour

正規化した中央絶対偏差

ロバストな標準偏差の推定値

MADN = \frac{median(|x_i - mean(x)|)}{0.674}

select 
    a.stn as stn,
    a.year as year,
    a.mo as month,
    a.da as day,
    a.temp as temp, 
    avg(a.temp) over (partition by a.stn) as temp_mu,
    stddev(a.temp) over (partition by a.stn) as temp_std,
    b.median as temp_median,
    b.madn as temp_madn
from [bigquery-public-data:noaa_gsod.gsod1929] as a
inner join 
(
    select 
        org.stn as stn,
        agg.median as median,
        nth(501, quantiles(abs(org.temp - agg.median), 1001))/0.675 as madn,
    from [bigquery-public-data:noaa_gsod.gsod1929] as org
    inner join 
    (
        select 
            stn,  nth(501, quantiles(temp, 1001)) as median
        from [bigquery-public-data:noaa_gsod.gsod1929]
        group by stn
    ) as agg on org.stn = agg.stn
    group by stn, median
) as b on a.stn = b.stn

Z値

BigQuery public datasetのNOAA Weather Dataから、ステーション毎に気温のZ値

select
	stn,
	year,
	mo,
	da,
	temp,
	(temp - mu)/std AS Z
from (
	select 
		stn,
		year,
		mo,
		da,
		temp, 
		avg(temp) over (partition by stn) as mu,
		stddev(temp) over (partition by stn) as std
	from [bigquery-public-data:noaa_gsod.gsod1929]
)
order by stn, year, mo, da