240 Z. Ma, J. Mi / Information Sciences 370–371 (2016) 239–255
the approximation operators, which are the important carriers of the rough set theory, and the brief properties characterize
the essential of the theory. Consequently, the axiomatizations of a variety of binary relations based generalized rough sets
have basically been solved [25,44,50–53,61] , as well as the axiomatizations of the coverings based generalized rough sets.
However, there is certain limitation in the extending of the above mentioned rough sets based on an arbitrary relation and
a covering by means of the classical definitions of approximation operators. For instance, when a relation is not a tolerance
relation, the approximation operators may lose many the basic properties of the classical operators. In addition, it is an
important issue for the correlation between the general relations and the coverings based generalized rough sets. And it is
not apparent to connect the two types of generalized rough sets for the reason that the coverings are more complex than
the partitions.
As one of the most important issues in rough set theory, uncertainty measures have been widely studied. The granula-
tion, as a powerful mechanism to evaluate uncertainties of the approximation spaces, was first introduced by Zadeh [58] . An
axiomatic definition of the knowledge granulation was proposed [18] , and a variety of forms of the knowledge granulation
were introduced in [8,11,17] . All these studies were dedicated to evaluating uncertainty of a set in terms of the partition
ability of a classification, and the measure methods are more intuitive and easily understood. To evaluate uncertainty of a
set, several numerical measures were presented by Pawlak [33] , which are accuracy and roughness of a set and approxi-
mation accuracy of a rough classification. Although the measures are effective in most cases, they have some restrictions in
some special cases. For instances, when the boundary (i.e., difference between upper and lower approximations) of some
sets are equal, it is not well to characterize the degree of closeness to lower or upper approximations of a set associated
with a kind of classification task.
In this paper, we first review several existing definitions of the lower and upper approximation operators, and show the
similarity and difference among them. In many situations, one always want to extend the classical rough sets to general-
ized rough sets but remain some basic properties. However, the approximation operators based on general binary relations
may not have the basic properties. At this point, our purpose is to avoid the limitation of the original definitions that are
straightly extended to the non-tolerance relations based generalized rough sets. Based on these considerations, we intro-
duce a measure of the predefined boundary regions based on the binary relations, and give a novel definitions of the lower
and upper approximation operators. We consider the properties of the novel operators and their axiomatic characteriza-
tions, and draw a comparison between the novel operators and the original ones. We also give the necessary and sufficient
conditions of their equipollence. Secondly, similar to the approach of the binary relations based generalized rough sets, the
novel definitions of coverings based approximation operators are proposed. Comparing the proposed definitions with the
existing ones, we point out that the covering based novel operators and the covering approximation operators proposed by
Zakowski [57] are equivalent. We also investigate the properties and axiomatic characterizations of the covering based novel
operators. Thirdly, we study the relationship between the covering and binary relation based generalized rough sets. In fact,
Zhu have established the relationship between a class of covering and a similarity relation based rough sets [62] . Bartol
et al.characterized the covering generated by a tolerance relation [2] . In our works, we show that the novel approximation
operators based on an arbitrary relation are equivalent to the ones based on a reduction of the corresponding covering.
Meanwhile, the novel operators based on a covering are equivalent to the ones based on a tolerance relation. In addition,
with the aid of the novel definitions, we also study relationship between the axiomatic characterization of the approxima-
tion operators based on the binary relations and the ones based on the coverings. Finally, we give the improved axiomatic
definition of the knowledge granulation, and introduce a more reasonable knowledge granulation based on the novel bound-
ary operators. It is more effective to characterize the approximation spaces based on a partition and a covering. For a object
set, to distinguish the different effect of the same boundaries on various sets, we also introduce some measure methods in
greater detail.
The construction of this paper is organized as follows. In Section 2 , we review and compare serval definitions of the
classical approximation operators, which are based on an equivalence relation. The novel definitions of the approximation
operators based on a relation are then presented, and their properties and axiomatic characterizations are also examined.
We also study novel definitions of a type of covering based approximation operators in Section 3 . In Section 4 , we combine
the two approximation operators based on the general binary relations and the coverings. Finally, we study the uncertain
measures in the approximation spaces in Section 5 . The conclusion and prospect of this paper are given in the last section.
2. The approximation operators based on binary relations via the boundary region
In this section, we discuss the general binary relations based approximation operators, and investigate the advantages
and disadvantages. To remove the defect of the approximation operators, we introduce a definition of boundary region
based on a relation and propose the novel definitions of the approximation operators by using the boundary region. And we
also study their properties and axiomatic characterizations.
2.1. The classical approximation operators
We recall basic definitions of the classical approximation operators and some other important concepts, such as neigh-
borhood, precision, precise set, etc. Then we review the basic properties of the approximation operators based on a binary
relation.