Mailing-List: contact dev-help@lucene.apache.org; run by ezmlm
Precedence: bulk
Reply-To: dev@lucene.apache.org
Date: Sat, 13 Feb 2016 17:22:18 +0000 (UTC)
From: "Joel Bernstein (JIRA)" <jira@apache.org>
To: dev@lucene.apache.org
Message-ID: <JIRA.12927355.1452027804000.46722.1455384138126@Atlassian.JIRA>
In-Reply-To: <JIRA.12927355.1452027804000@Atlassian.JIRA>
References: <JIRA.12927355.1452027804000@Atlassian.JIRA>
 <JIRA.12927355.1452027804581@arcas>
Subject: [jira] [Commented] (SOLR-8492) Add LogisticRegressionQuery and
 LogitStream
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: quoted-printable


    [ https://issues.apache.org/jira/browse/SOLR-8492?page=3Dcom.atlassian.=
jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=3D15146=
074#comment-15146074 ]=20

Joel Bernstein commented on SOLR-8492:
--------------------------------------

[~caomanhdat], I've been testing out the algorithm by comparing it to R. I'=
ve attached the data set I'm using to the ticket. I'm attempting make sense=
 of the results but they don't seem to match up.

I ran the following LogitStream call:
{code}
logit(collection3, q=3D"*:*", features=3D"married_i", outcome=3D"buy_i", ma=
xIterations=3D100)
{code}
And got this for the final weights:
{code}
{"features":["married_i"],"weights":[-0.5048080969966502],"error":0.3808560=
947044004}
{code}

In R, here is the output for the same data set. *Note* that *Is.Married* is=
 the same field as *married_i* in Solr and same for *Buy* and *buy_i*

{code}
Call:
glm(formula =3D Buy ~ Is.Married, family =3D "binomial", data =3D mydata)

Deviance Residuals:=20
    Min       1Q   Median       3Q      Max =20
-0.9687  -0.4201  -0.4201  -0.4201   2.2232 =20

Coefficients:
            Estimate Std. Error z value Pr(>|z|)   =20
(Intercept)  -2.3830     0.1718 -13.870   <2e-16 ***
Is.Married    1.8699     0.2184   8.563   <2e-16 ***
---
Signif. codes:  0 =E2=80=98***=E2=80=99 0.001 =E2=80=98**=E2=80=99 0.01 =E2=
=80=98*=E2=80=99 0.05 =E2=80=98.=E2=80=99 0.1 =E2=80=98 =E2=80=99 1

(Dispersion parameter for binomial family taken to be 1)

    Null deviance: 646.05  on 672  degrees of freedom
Residual deviance: 564.47  on 671  degrees of freedom
AIC: 568.47

Number of Fisher Scoring iterations: 5

{code}

It appears that they are giving different results. Am I performing the test=
 correctly?


> Add LogisticRegressionQuery and LogitStream
> -------------------------------------------
>
>                 Key: SOLR-8492
>                 URL: https://issues.apache.org/jira/browse/SOLR-8492
>             Project: Solr
>          Issue Type: New Feature
>            Reporter: Joel Bernstein
>         Attachments: SOLR-8492.patch, SOLR-8492.patch, SOLR-8492.patch, S=
OLR-8492.patch, SOLR-8492.patch, SOLR-8492.patch
>
>
> This ticket is to add a new query called a LogisticRegressionQuery (LRQ).
> The LRQ extends AnalyticsQuery (http://joelsolr.blogspot.com/2015/12/unde=
rstanding-solrs-analyticsquery.html) and returns a DelegatingCollector that=
 implements a Stochastic Gradient Descent (SGD) optimizer for Logistic Regr=
ession.
> This ticket also adds the LogitStream which leverages Streaming Expressio=
ns to provide iteration over the shards. Each call to LogitStream.read() ca=
lls down to the shards and executes the LogisticRegressionQuery. The model =
data is collected from the shards and the weights are averaged and sent bac=
k to the shards with the next iteration. Each call to read() returns a Tupl=
e with the averaged weights and error from the shards. With this approach t=
he LogitStream streams the changing model back to the client after each ite=
ration.
> The LogitStream will return the EOF Tuple when it reaches the defined max=
Iterations. When sent as a Streaming Expression to the Stream handler this =
provides parallel iterative behavior. This same approach can be used to imp=
lement other parallel iterative algorithms.
> The initial patch has  a test which simply tests the mechanics of the ite=
ration. More work will need to be done to ensure the SGD is properly implem=
ented. The distributed approach of the SGD will also need to be reviewed. =
=20
> This implementation is designed for use cases with a small number of feat=
ures because each feature is it's own discreet field.
> An implementation which supports a higher number of features would be pos=
sible by packing features into a byte array and storing as binary DocValues=
.
> This implementation is designed to support a large sample set. With a lar=
ge number of shards, a sample set into the billions may be possible.
> sample Streaming Expression Syntax:
> {code}
> logit(collection1, features=3D"a,b,c,d,e,f" outcome=3D"x" maxIterations=
=3D"80")
> {code}


--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: dev-help@lucene.apache.org