Re: How to classify unseen multi instance data?

> > Even for multi-instance data you only use a single Instance object,
> > representing a single bag, for making predictions.
>
> My confusion stemmed from the somewhat unfortunate terminology, whereby
> I thought an Instance object is single-instance and Instances is
> multi-instance - which I now believe to be wrong. An Instance can either
> be single- or multi-instance and if it is multi-instance we also call it
> a bag. Instances is then either:
>
> - a bunch of bags or
> - a bunch of single-instance Instance objects

Instances = dataset, a collection of examples
Instance = single example

> The next point to note is that training a model is done off an Instances
> object, whereas classification of unseen data is using an Instance.
> Therefore the answer to my original question appears to be thus:
>
> When training a model, first get the data from somewhere (ARFF file,
> database, ...) and then convert it to an Instances object. When
> classifying unseen data, use the exact same code to create an Instances
> object (optionally skipping the class values, which are not needed
> here). Then it's just a matter of passing Instances.firstInstance() to
> the classifier. Hopefully I explained it correctly now :)

At prediction time you re-use (rather than re-create) the header from
the training data to create new Instance objects. These Instance
objects you then present to the classifier to obtain predictions.

> FWIW, from a software engineering standpoint single- and multi-instance
> data should be distinguishable by their type (Java class).
> Multi-instance data should probably implement an additional interface.

Yes and no. Whilst it would be nice to determine this via the Java
type, the general RELATIONAL attribute is column-based and not
row-based (and we already have DenseInstance and SparseInstance row
types). The idea behind the attribute is to represent data from a
relational database. But, this attribute can also be used to represent
the simpler multi-instance data, by encapsulating the bag in such an
attribute.

Originally, multi-instance learning was available through a separate
framework called MILK. Eibe then added the RELATIONAL attribute type
to Weka, to make it possible having multi-instance learning as part of
Weka via optional Weka packages.

The RELAGGS filter is specifically designed for processing data from a
relational database (the original code worked straight off JDBC
databases):
https://weka.sourceforge.io/doc.stable/weka/filters/unsupervised/attribute/RELAGGS.html

So, using the attribute-based approach rather than the Java-type-row
approach is more flexible. :-)

Cheers, Peter
-- 
Peter Reutemann
Dept. of Computer Science
University of Waikato, Hamilton, NZ
Mobile +64 22 190 2375
https://www.cs.waikato.ac.nz/~fracpete/
http://www.data-mining.co.nz/
_______________________________________________
Wekalist mailing list -- [email protected]
Send posts to [email protected]
To unsubscribe send an email to [email protected]
To subscribe, unsubscribe, etc., visit https://list.waikato.ac.nz/postorius/lists/wekalist.list.waikato.ac.nz
List etiquette: http://www.cs.waikato.ac.nz/~ml/weka/mailinglist_etiquette.html