This example explains how to run ensemble learning in Hivemall.
Two heads are better than one? Let's verify it by ensemble learning.

[Case1] Model ensemble/mixing

training

SET hive.exec.parallel=true;
SET hive.exec.parallel.thread.number=8;
SET mapred.reduce.tasks=4;

drop table news20mc_ensemble_model1;
create table news20mc_ensemble_model1 as
select 
 label, 
 -- cast(feature as int) as feature, -- hivemall v0.1
 argmin_kld(feature, covar) as feature, -- hivemall v0.2 or later
 voted_avg(weight) as weight
from 
 (select 
     -- train_multiclass_cw(add_bias(features),label) as (label,feature,weight)      -- hivemall v0.1
     train_multiclass_cw(add_bias(features),label) as (label,feature,weight,covar)   -- hivemall v0.2 or later
  from 
     news20mc_train_x3
  union all
  select 
     -- train_multiclass_arow(add_bias(features),label) as (label,feature,weight)    -- hivemall v0.1
     train_multiclass_arow(add_bias(features),label) as (label,feature,weight,covar) -- hivemall v0.2 or later
  from 
     news20mc_train_x3
  union all
  select 
     -- train_multiclass_scw(add_bias(features),label) as (label,feature,weight)     -- hivemall v0.1
     train_multiclass_scw(add_bias(features),label) as (label,feature,weight,covar)  -- hivemall v0.2 or later
  from 
     news20mc_train_x3
 ) t 
group by label, feature;

-- reset to the default
SET hive.exec.parallel=false;
SET mapred.reduce.tasks=-1;

prediction

create or replace view news20mc_ensemble_predict1 
as
select 
  rowid, 
  m.col0 as score, 
  m.col1 as label
from (
select
   rowid, 
   maxrow(score, label) as m
from (
  select
    t.rowid,
    m.label,
    sum(m.weight * t.value) as score
  from 
    news20mc_test_exploded t LEFT OUTER JOIN
    news20mc_ensemble_model1 m ON (t.feature = m.feature)
  group by
    t.rowid, m.label
) t1
group by rowid
) t2;

evaluation

create or replace view news20mc_ensemble_submit1 as
select 
  t.label as actual, 
  pd.label as predicted
from 
  news20mc_test t JOIN news20mc_ensemble_predict1 pd 
    on (t.rowid = pd.rowid);
select count(1)/3993 from news20mc_ensemble_submit1 
where actual == predicted;

0.8494866015527173

Unfortunately, too many cooks spoil the broth in this case :-(

Algorithm Accuracy
AROW 0.8474830954169797
SCW2 0.8482344102178813
Ensemble(model) 0.8494866015527173
CW 0.850488354620586

[Case2] Prediction ensemble

prediction

create or replace view news20mc_pred_ensemble_predict1 
as
select 
  rowid, 
  m.col1 as label
from (
  select
    rowid, 
    maxrow(cnt, label) as m
  from (
    select
      rowid,
      label,
      count(1) as cnt
    from (
      select * from news20mc_arow_predict1
      union all
      select * from news20mc_scw2_predict1
      union all
      select * from news20mc_cw_predict1
    ) t1
    group by rowid, label
  ) t2
  group by rowid
) t3;

evaluation

create or replace view news20mc_pred_ensemble_submit1 as
select 
  t.label as actual, 
  pd.label as predicted
from 
  news20mc_test t JOIN news20mc_pred_ensemble_predict1 pd 
    on (t.rowid = pd.rowid);
select count(1)/3993 from news20mc_pred_ensemble_submit1 
where actual == predicted;

0.8499874780866516

Unfortunately, too many cooks spoil the broth in this case too :-(

Algorithm Accuracy
AROW 0.8474830954169797
SCW2 0.8482344102178813
Ensemble(model) 0.8494866015527173
Ensemble(prediction) 0.8499874780866516
CW 0.850488354620586

results matching ""

    No results matching ""