a9a binary classification (logistic regression)

a9a

http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/binary.html#a9a

Training with iterations is OBSOLUTE in Hivemall.
Using amplifier and shuffling inputs is RECOMMENDED in Hivemall.

---

UDF preparation

select count(1) from a9atrain;
-- set total_steps ideally be "count(1) / #map tasks"
set hivevar:total_steps=32561;

select count(1) from a9atest;
set hivevar:num_test_instances=16281;

training

create table a9a_model1 
as
select 
 cast(feature as int) as feature,
 avg(weight) as weight
from 
 (select 
     logress(addBias(features),label,"-total_steps ${total_steps}") as (feature,weight)
  from 
     a9atrain
 ) t 
group by feature;

"-total_steps" option is optional for logress() function.
I recommend you NOT to use options (e.g., total_steps and eta0) if you are not familiar with those options. Hivemall then uses an autonomic ETA (learning rate) estimator.

prediction

create or replace view a9a_predict1 
as
WITH a9atest_exploded as (
select 
  rowid,
  label,
  extract_feature(feature) as feature,
  extract_weight(feature) as value
from 
  a9atest LATERAL VIEW explode(addBias(features)) t AS feature
)
select
  t.rowid, 
  sigmoid(sum(m.weight * t.value)) as prob,
  CAST((case when sigmoid(sum(m.weight * t.value)) >= 0.5 then 1.0 else 0.0 end) as FLOAT) as label
from 
  a9atest_exploded t LEFT OUTER JOIN
  a9a_model1 m ON (t.feature = m.feature)
group by
  t.rowid;

evaluation

create or replace view a9a_submit1 as
select 
  t.label as actual, 
  pd.label as predicted, 
  pd.prob as probability
from 
  a9atest t JOIN a9a_predict1 pd 
    on (t.rowid = pd.rowid);

select count(1) / ${num_test_instances} from a9a_submit1 
where actual == predicted;

0.8430071862907684