[prev in list] [next in list] [prev in thread] [next in thread] 

List:       wekalist
Subject:    Re: [Wekalist] predicting a low frequency class value
From:       Eibe Frank <eibe.frank () waikato ! ac ! nz>
Date:       2018-05-26 0:49:44
Message-ID: 43c7424ae8b0cb4337c729452bb90312 () mail ! gmail ! com
[Download RAW message or body]

In CostSensitiveClassifier, try switching to the minimum expected cost
approach (there is a flag for that). It often works well, but requires a
base classifier that produces good class probability estimates. Bagging
can often provide those. Try that and let us know how you go. You may also
want to try turning pruning off in the decision tree learner that is used.

Another option is to try to augment the data in the minority class. The
SMOTE method is a widely-used approach for doing this. It is implemented
in the SMOTE filter. Use it in conjunction with the FilteredClassifier and
a base classifier of your choice.

Cheers,
Eibe

> -----Original Message-----
> From: wekalist-bounces@list.waikato.ac.nz <wekalist-
> bounces@list.waikato.ac.nz> On Behalf Of Lindsay Peters
> Sent: Thursday, 24 May 2018 10:30 PM
> To: 'Weka machine learning workbench list.'
<wekalist@list.waikato.ac.nz>
> Subject: [Wekalist] predicting a low frequency class value
>
> Hi Eibe,
>
> The domain I'm trying to analyse has a binary-valued class, and the "no"
value
> is infrequently seen, i.e. 1% to perhaps 10% of cases.
>
> The goal however is actually to predict instances giving this "no"
value, and
> the consequence of a false negative is high, much higher than a false
positive.
>
> I've tried using weka J48 with a 2x2 cost matrix weighted for the false
> negatives, but the results are very poor.
>
> Can you please point me to classifiers which may be better suited to
this
> domain, or other techniques?
>
> Thanks in advance,
>
> Lindsay Peters
>
> _______________________________________________
> Wekalist mailing list
> Send posts to: Wekalist@list.waikato.ac.nz List info and subscription
status:
> https://list.waikato.ac.nz/mailman/listinfo/wekalist
> List etiquette:
> http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html
_______________________________________________
Wekalist mailing list
Send posts to: Wekalist@list.waikato.ac.nz
List info and subscription status: https://list.waikato.ac.nz/mailman/listinfo/wekalist
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html
[prev in list] [next in list] [prev in thread] [next in thread]