| import polars as pl, numpy as np |
| from sklearn.metrics import roc_auc_score |
| def wauc(df,c,y='clicked',g='eid'): |
| d=df.select(g,y,pl.col(c).cast(pl.Float64).fill_null(-1e18).alias('s')).with_columns(pl.col('s').rank('average').over(g).alias('rk'), |
| pl.col(y).sum().over(g).alias('P'),pl.len().over(g).alias('N')).filter((pl.col('P')>0)&(pl.col('P')<pl.col('N'))) |
| e=d.filter(pl.col(y)==1).group_by(g).agg(((pl.col('rk').sum()-pl.col('P').first()*(pl.col('P').first()+1)/2)/(pl.col('P').first()*(pl.col('N').first()-pl.col('P').first()))).alias('a')) |
| return float(e['a'].mean()) |
| def gauc(df,c,y='clicked'): |
| d=df.select(y,pl.col(c).cast(pl.Float64)).drop_nulls().drop_nans() |
| return float(roc_auc_score(d[y].to_numpy(),d[c].to_numpy())) |
| def eauc(df,c,y='clicked',g='eid'): |
| e=df.group_by(g).agg(pl.col(y).max(),pl.col(c).first()).drop_nulls() |
| return float(roc_auc_score(e[y].to_numpy(),e[c].cast(pl.Float64).to_numpy())) |
| def rep(df,c,event_level=False): |
| out={} |
| for v in ('v25','v26'): |
| d=df.filter(pl.col(v)==1) |
| a=gauc(d,c); out[v]=dict(glob=round(max(a,1-a),4),dirn='+' if a>=.5 else '-') |
| if event_level: e=eauc(d,c); out[v]['event']=round(max(e,1-e),4) |
| else: out[v]['within']=round(wauc(d,c),4) |
| return out |
|
|
