184 Applied Data Mining
to decide the relevant labels and irrelevant labels. For label l
i
, it is predicted
as a relevant label if f
i
(x) ≥ t, otherwise it is predicted as an irrelevant label.
LR is a very important issue in multi-label learning, not only it is an effective
means of solving multi-label problems, but also it’s more suitable for some
real applications. For example, users of tag recommendation system might
prefer seeing the most interesting tags top the tag list, instead of just a set
of tags. So it poses a very interesting and useful way to solving multi-
label classifi cation. Usually, LR models can also be learned for single-label
instances to solve the multi-class problems.
Nowadays, researchers have proposed lots of effective and effi cient
methods for multi-label classification, and they mainly fall into two
main categories: (1) algorithm adaptation (2) problem transformation [9].
Algorithm adaptation methods extend traditional single-label algorithms,
such as kNN, decision tree, Naive Bayes etc., in order to enable them
to handle multi-label data directly [4, 10, 11]. By contrast, problem
transformation methods decompose the multi-label instances into one or
several single-label instances, thus existing methods could be used without
modifi cation [12, 13, 14]. In other words, the algorithm adaptation strategy
is to fi t the algorithms to data, whereas the problem transformation strategy
is to fi t data to the algorithms. The primary difference between these two
strategies is that the algorithm adaptation is algorithm-specifi c, therefore
the strategy used in one method can not be applied to another one usually.
Nevertheless the problem transformation is algorithm-independent, so it
is more fl exible and can be used with any existing models. The following
sections will elaborate on them by different categories.
8.3 Problem Transformation
As mentioned above, problem transformation is a fundamental strategy for
tackling multi-label problems. It enables most of the existing methods to
work easily, while making few modifi cations to them. Hence it gets much
popularity among researchers, and various methods based on it have
been proposed [12, 13, 14, 15, 16]. Several simple methods of this kind are
All Label Assignment (ALA), No Label Assignment (NLA), Largest Label
Assignment (LLA) and Smallest Label Assignment (SLA), as summarized
by Chen et al. and used for multi-label document transformation [17]. Let’s
explain these methods through a multi-label dataset shown in Table 8.1.