Confidece Weighted (CW)

training

drop table news20b_cw_model1;
create table news20b_cw_model1 as
select 
 feature,
 -- voted_avg(weight) as weight -- [hivemall v0.1]
 argmin_kld(weight, covar) as weight -- [hivemall v0.2 or later]
from 
 (select 
     -- train_cw(add_bias(features), label) as (feature, weight) -- [hivemall v0.1]
     train_cw(add_bias(features), label) as (feature, weight, covar) -- [hivemall v0.2 or later]
  from 
     news20b_train_x3
 ) t 
group by feature;

prediction

create or replace view news20b_cw_predict1 
as
select
  t.rowid, 
  sum(m.weight * t.value) as total_weight,
  case when sum(m.weight * t.value) > 0.0 then 1 else -1 end as label
from 
  news20b_test_exploded t LEFT OUTER JOIN
  news20b_cw_model1 m ON (t.feature = m.feature)
group by
  t.rowid;

evaluation

WITH submit as (
select 
  t.rowid,
  t.label as actual, 
  p.label as predicted
from 
  news20b_test t 
  JOIN news20b_cw_predict1 p
    on (t.rowid = p.rowid)
)
select sum(if(actual = predicted, 1, 0)) / count(1) as accuracy
from submit;

0.9655724579663731

Adaptive Regularization of Weight Vectors (AROW)

training

drop table news20b_arow_model1;
create table news20b_arow_model1 as
select 
 feature,
 -- voted_avg(weight) as weight -- [hivemall v0.1]
 argmin_kld(weight, covar) as weight -- [hivemall v0.2 or later]
from 
 (select 
     -- train_arow(add_bias(features),label) as (feature,weight) -- [hivemall v0.1]
     train_arow(add_bias(features),label) as (feature,weight,covar) -- [hivemall v0.2 or later]
  from 
     news20b_train_x3
 ) t 
group by feature;

prediction

create or replace view news20b_arow_predict1 
as
select
  t.rowid, 
  sum(m.weight * t.value) as total_weight,
  case when sum(m.weight * t.value) > 0.0 then 1 else -1 end as label
from 
  news20b_test_exploded t LEFT OUTER JOIN
  news20b_arow_model1 m ON (t.feature = m.feature)
group by
  t.rowid;

evaluation

WITH submit as (
select
  t.rowid, 
  t.label as actual, 
  p.label as predicted
from 
  news20b_test t
  JOIN news20b_arow_predict1 p
    on (t.rowid = p.rowid)
)
select sum(if(actual = predicted, 1, 0)) / count(1) as accuracy
from submit;

0.9659727782225781

Soft Confidence-Weighted (SCW1)

training

drop table news20b_scw_model1;
create table news20b_scw_model1 as
select 
 feature,
 -- voted_avg(weight) as weight -- [hivemall v0.1]
 argmin_kld(weight, covar) as weight -- [hivemall v0.2 or later]
from 
 (select 
     -- train_scw(add_bias(features),label) as (feature,weight) -- [hivemall v0.1]
     train_scw(add_bias(features),label) as (feature,weight,covar) -- [hivemall v0.2 or later]
  from 
     news20b_train_x3
 ) t 
group by feature;

prediction

create or replace view news20b_scw_predict1 
as
select
  t.rowid, 
  sum(m.weight * t.value) as total_weight,
  case when sum(m.weight * t.value) > 0.0 then 1 else -1 end as label
from 
  news20b_test_exploded t LEFT OUTER JOIN
  news20b_scw_model1 m ON (t.feature = m.feature)
group by
  t.rowid;

evaluation

WITH submit as (
  select 
    t.rowid, 
    t.label as actual, 
    p.label as predicted
  from 
    news20b_test t JOIN news20b_scw_predict1 p
      on (t.rowid = p.rowid)
)
select sum(if(actual = predicted, 1, 0)) / count(1) as accuracy
from submit

0.9661729383506805

Soft Confidence-Weighted (SCW2)

training

drop table news20b_scw2_model1;
create table news20b_scw2_model1 as
select 
 feature,
 -- voted_avg(weight) as weight -- [hivemall v0.1]
 argmin_kld(weight, covar) as weight -- [hivemall v0.2 or later]
from 
 (select 
     -- train_scw2(add_bias(features),label) as (feature,weight)    -- [hivemall v0.1]
     train_scw2(add_bias(features),label) as (feature,weight,covar) -- [hivemall v0.2 or later]
  from 
     news20b_train_x3
 ) t 
group by feature;

prediction

create or replace view news20b_scw2_predict1 
as
select
  t.rowid, 
  sum(m.weight * t.value) as total_weight,
  case when sum(m.weight * t.value) > 0.0 then 1 else -1 end as label
from 
  news20b_test_exploded t LEFT OUTER JOIN
  news20b_scw2_model1 m ON (t.feature = m.feature)
group by
  t.rowid;

evaluation

WITH submit as (
select 
  t.rowid, 
  t.label as actual, 
  pd.label as predicted
from 
  news20b_test t
  JOIN news20b_scw2_predict1 pd 
    on (t.rowid = pd.rowid)
)
select sum(if(actual = predicted, 1, 0)) / count(1) as accuracy
from submit;

0.9579663730984788

--

My recommendation is AROW for classification.