[jira] [Updated] (MADLIB-1517) Do not include intercept as predictor for PMML exported by the regression models

["Nikhil Kak (Jira)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/MADLIB-1517?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Nikhil Kak updated MADLIB-1517:
-------------------------------
    Summary: Do not include intercept as predictor for PMML exported by the regression models  (was: Do not include intercept as predictor for regression models)

> Do not include intercept as predictor for PMML exported by the regression models
> --------------------------------------------------------------------------------
>
>                 Key: MADLIB-1517
>                 URL: https://issues.apache.org/jira/browse/MADLIB-1517
>             Project: Apache MADlib
>          Issue Type: Bug
>          Components: Module: Utilities
>            Reporter: Ekta Khanna
>            Priority: Major
>             Fix For: v2.2.0
>
>
> Our regression models include the intercept as a predictor in the exported pmml file which isn't ideal when using this pmml file for predictions
> h1. Reproduction Steps
> Create glm model and generate pmml
> {code:java}
> CREATE TABLE warpbreaks(
>     id      serial,
>     breaks  integer,
>     wool    char(1),
>     tension char(1)
> );
> INSERT INTO warpbreaks(breaks, wool, tension) VALUES
> (26, 'A', 'L'),
> (30, 'A', 'L'),
> (54, 'A', 'L'),
> (25, 'A', 'L'),
> (70, 'A', 'L'),
> (52, 'A', 'L'),
> (51, 'A', 'L'),
> (26, 'A', 'L'),
> (67, 'A', 'L'),
> (18, 'A', 'M'),
> (21, 'A', 'M'),
> (29, 'A', 'M'),
> (17, 'A', 'M'),
> (12, 'A', 'M'),
> (18, 'A', 'M'),
> (35, 'A', 'M'),
> (30, 'A', 'M'),
> (36, 'A', 'M'),
> (36, 'A', 'H'),
> (21, 'A', 'H'),
> (24, 'A', 'H'),
> (18, 'A', 'H'),
> (10, 'A', 'H'),
> (43, 'A', 'H'),
> (28, 'A', 'H'),
> (15, 'A', 'H'),
> (26, 'A', 'H'),
> (27, 'B', 'L'),
> (14, 'B', 'L'),
> (29, 'B', 'L'),
> (19, 'B', 'L'),
> (29, 'B', 'L'),
> (31, 'B', 'L'),
> (41, 'B', 'L'),
> (20, 'B', 'L'),
> (44, 'B', 'L'),
> (42, 'B', 'M'),
> (26, 'B', 'M'),
> (19, 'B', 'M'),
> (16, 'B', 'M'),
> (39, 'B', 'M'),
> (28, 'B', 'M'),
> (21, 'B', 'M'),
> (39, 'B', 'M'),
> (29, 'B', 'M'),
> (20, 'B', 'H'),
> (21, 'B', 'H'),
> (24, 'B', 'H'),
> (17, 'B', 'H'),
> (13, 'B', 'H'),
> (15, 'B', 'H'),
> (15, 'B', 'H'),
> (16, 'B', 'H'),
> (28, 'B', 'H');
> SELECT create_indicator_variables('warpbreaks', 'warpbreaks_dummy', 'wool,tension');
> DROP TABLE IF EXISTS glm_model, glm_model_summary;
> SELECT glm('warpbreaks_dummy',
>            'glm_model',
>            'breaks',
>            'ARRAY[1.0,"wool_B","tension_M", "tension_H"]',
>            'family=poisson, link=log');
> COPY (SELECT madlib.pmml('glm_model')) TO '/tmp/glm.pmml';
> SELECT madlib.glm_predict(coef,  ARRAY[1, 0, 1, 0]::float8[], 'log') FROM glm_model;
>     glm_predict
> --------------------
>  29.097222222222218
> SELECT madlib.glm_predict(coef, ARRAY[1, 0, 0, 0]::float8[], 'log') FROM glm_model;
>     glm_predict
> --------------------
>  40.123538011695906
> {code}
> Use pypmml to predict the data using the generated pmml:
> {code:java}
> from pypmml import Model
> model = Model.fromFile('/tmp/glm.pmml')
> data = [{ "wool_B": 0, "tension_M": 0, "tension_H": 0},
>         { "wool_B": 0, "tension_M": 1, "tension_H": 0},
>         ]
> for d in data:
>     result = model.predict(d)
>     print(d)
>     print(result)
> {'wool_B': 0, 'tension_M': 0, 'tension_H': 0}
> {'predicted_breaks_pmml_prediction': nan}
> {'wool_B': 0, 'tension_M': 1, 'tension_H': 0}
> {'predicted_breaks_pmml_prediction': nan}
> {code}
> Obviously the nan results are wrong. To make it work with the existing pmml file, we need to modify the input a bit.
> {code:java}
> from pypmml import Model
> model = Model.fromFile('/tmp/glm.pmml')
> data = [{ "1.0": 1, "wool_B": 0, "tension_M": 0, "tension_H": 0},
>         { "1.0": 1, "wool_B": 0, "tension_M": 1, "tension_H": 0}
>         ]
> for d in data:
>     result = model.predict(d)
>     print(d)
>     print(result)
> {'1.0': 1, 'wool_B': 0, 'tension_M': 0, 'tension_H': 0}
> {'predicted_breaks_pmml_prediction': 40.123538011695906}
> {'1.0': 1, 'wool_B': 0, 'tension_M': 1, 'tension_H': 0}
> {'predicted_breaks_pmml_prediction': 29.097222222222218}
> {code}
> Now these values match the madlib glm_predict output
> h1. Goal
> The goal of this story is to fix the pmml code so that we don't need to awkwardly pass the intercept as "'1.0':1" before the beginning of each data row. Note the the name "1.0" is used because that's how it's stored in the pmml file



--
This message was sent by Atlassian Jira
(v8.20.10#820010)