Rough Neighborhood Set (NRS) Overview

• Rough Neighborhood Set (NRS) is a generalization of rough sets that replaces equivalence classes with local neighborhoods derived from coverings, metrics, or similarities.
• The model employs neighborhood lower and upper approximations to capture certainty and possibility, with applications in covering-based and metric data approximations.
• Extensions like granular-ball rough sets unify discrete and continuous data handling, enhancing rule induction techniques such as Neighborhood-LERS for improved classification.

Searching arXiv for recent and foundational papers on neighborhood rough sets and closely related formulations. arxiv_search(query="neighborhood rough set rough set covering-based rough sets partition neighborhoods", max_results=10) Rough Neighborhood Set (NRS) denotes a rough-set model in which the basic granules are neighborhoods rather than equivalence classes. In the most general formulation used in the rough-set literature, an NRS is based on a neighborhood approximation space $(U,N)$, where $U$ is a nonempty universe and $N(x)\subseteq U$ is a nonempty neighborhood assigned to each object $x\in U$. For any target set $X\subseteq U$, the neighborhood lower approximation collects objects whose neighborhoods are entirely contained in $X$, while the neighborhood upper approximation collects objects whose neighborhoods intersect $X$. In this way, NRS generalizes classical Pawlak rough sets by replacing partition-based granulation with local granules induced by coverings, relations, metrics, similarities, or graph neighborhoods (Fujita et al., 7 Apr 2026, Yao et al., 2012).

1. Formal framework and approximation semantics

In the handbook treatment of neighborhood rough sets, a neighborhood approximation space is written as $(U,N)$, with $N:U\to\mathcal{P}(U)$. The defining approximations of a set $X\subseteq U$ are

$U$0

The associated neighborhood rough set of $U$1 is the pair

$U$2

The standard derived regions are

$U$3

These definitions preserve the usual certainty/possibility semantics of rough sets: the lower approximation contains objects definitely in $U$4 under the adopted granulation, and the upper approximation contains objects possibly in $U$5 (Fujita et al., 7 Apr 2026).

A central feature of NRS is that the granulation need not be induced by an equivalence relation. This permits overlapping neighborhoods, non-partitional local structure, and direct use of similarity or distance information. At the same time, the Pawlak model is recovered as a special case: if one sets $U$6 for an equivalence relation $U$7, then the neighborhood lower and upper approximations coincide with the classical rough approximations $U$8 and $U$9 (Fujita et al., 7 Apr 2026).

Under the assumptions typically used in the handbook exposition, NRS retains basic properties such as

$N(x)\subseteq U$0

together with monotonicity in the target set. This places neighborhood rough sets within the broader family of approximation systems that model uncertainty through lower and upper bounds induced by data granulation rather than by probabilistic assumptions (Fujita et al., 7 Apr 2026).

2. Covering-induced neighborhoods and partition conditions

An important formulation of NRS arises in covering-based rough sets. Here one begins with a finite, nonempty universe $N(x)\subseteq U$1 and a covering $N(x)\subseteq U$2, that is, a family of nonempty subsets of $N(x)\subseteq U$3 satisfying

$N(x)\subseteq U$4

Each member $N(x)\subseteq U$5 is a covering block. The neighborhood induced by the covering is defined by

$N(x)\subseteq U$6

and the induced neighborhood family is

$N(x)\subseteq U$7

The neighborhood rough approximations then take the now-standard form

$N(x)\subseteq U$8

In this setting, an NRS may be viewed either as $N(x)\subseteq U$9 or as $x\in U$0 with neighborhoods induced by the covering (Yao et al., 2012).

The structure of these induced neighborhoods is constrained. A basic proposition states that if $x\in U$1, then $x\in U$2; moreover, if $x\in U$3 and $x\in U$4, then $x\in U$5. This observation underlies the question studied in depth in the covering-based literature: when does the family $x\in U$6 itself form a partition of $x\in U$7? The problem matters because, when the induced neighborhoods are pairwise disjoint and exhaustive, neighborhood-based approximations behave much more like classical Pawlak approximations (Yao et al., 2012).

The covering-based analysis introduces several structural notions. A covering block $x\in U$8 is reducible if it can be expressed as the union of other blocks, and the reduct $x\in U$9 is the set of irreducible blocks. A crucial proposition is

$X\subseteq U$0

so reducible blocks do not affect neighborhoods. The paper also defines the membership repeat degree

$X\subseteq U$1

the common block repeat degree

$X\subseteq U$2

and the $X\subseteq U$3-excluded number of $X\subseteq U$4,

$X\subseteq U$5

A key lemma establishes

$X\subseteq U$6

Using these quantities, the paper gives a necessary and sufficient condition for $X\subseteq U$7 to form a partition: $X\subseteq U$8 Equivalently, for any pair $X\subseteq U$9, either each lies in the other’s neighborhood or neither does. The same work also gives two sufficient conditions: if $X$0 is a partition, then $X$1 forms a partition; and if all covering blocks are uniform blocks, meaning all elements in a block have the same repeat degree, then $X$2 forms a partition (Yao et al., 2012).

When the partition condition holds, neighborhood rough approximations collapse to Pawlak-style approximations with respect to the partition $X$3. This supplies a precise bridge between covering-based NRS and equivalence-based rough sets (Yao et al., 2012).

3. Metric neighborhoods, continuous attributes, and numeric information systems

A second major line of development defines neighborhoods directly in metric spaces. In the summarization work on supervised multi-document text summarization, the starting point is a numeric information system

$X$4

where $X$5 is the universe of objects and $X$6 is a set of attributes. For an object $X$7, the value of $X$8 on attribute $X$9 is written $X$0. The paper uses the Minkowski distance

$X$1

with Manhattan distance for $X$2, Euclidean distance for $X$3, and Chebyshev distance for $X$4. The corresponding $X$5-neighborhood of $X$6 is

$X$7

Lower and upper approximations are then

$X$8

with boundary and negative regions defined analogously from these neighborhood-based approximations (Yadav, 2021).

This metric formulation is motivated by the fact that many application domains are described by continuous or primarily numeric attributes. The summarization paper states explicitly that a key drawback of classical LERS is the need for nominal attributes, whereas the computed sentence attributes are primarily numeric. The unified granular-ball paper states the same issue more generally for Pawlak rough sets: exact-value indiscernibility is poorly matched to continuous data, often forcing discretization and its associated information loss. Neighborhood rough sets address this by defining granules through distance rather than strict equivalence (Yadav, 2021, Xia et al., 2022).

The same general picture appears in the handbook taxonomy, which treats neighborhood models as “metric/graph/kNN style, local granules.” It records two canonical constructions: relation-induced neighborhoods,

$X$9

and metric-based neighborhoods,

$(U,N)$0

Tolerance-based and similarity-threshold models fit this scheme as special cases. In the handbook’s formulation, neighborhood rough sets are therefore not a single algorithm but a family of approximation systems parameterized by the way neighborhoods are generated (Fujita et al., 7 Apr 2026).

The literature summarized here also records specific limitations of classical NRS. The granular-ball paper emphasizes that standard neighborhood rough sets lose the ability to use equivalence classes for knowledge representation, require a neighborhood radius $(U,N)$1, and can suffer from “heterogeneous transmission,” where overlapping neighborhoods from different decision classes undermine clean granulation (Xia et al., 2022).

4. Rule induction and supervised learning: the Neighborhood-LERS model

A concrete algorithmic use of NRS appears in supervised multi-document text summarization through Neighborhood-LERS (NLER), an improved LERS algorithm in which rough approximations and coverings are computed by neighborhood rough sets rather than by classical rough-set equivalence classes (Yadav, 2021).

The workflow described in that paper begins by computing a decision attribute $(U,N)$2 for training sentences using ROUGE scores relative to reference summaries. Sentences are then parsed and represented by numeric features $(U,N)$3, yielding an information system $(U,N)$4 with decision attribute $(U,N)$5. The algorithm next determines a global covering $(U,N)$6 “using Neighborhood Rough Sets based distance for each feature,” with each feature placed in an $(U,N)$7-neighborhood for a specific $(U,N)$8. The experiments explicitly use radii such as $(U,N)$9 and $N:U\to\mathcal{P}(U)$0, denoted $N:U\to\mathcal{P}(U)$1 and $N:U\to\mathcal{P}(U)$2 (Yadav, 2021).

Rule generation proceeds from neighborhood-based approximations of the decision classes. Sentences whose neighborhoods lie entirely within a decision class belong to its lower approximation and act as certain examples; sentences whose neighborhoods overlap a class belong to its upper approximation. The algorithm then generates rules from the global covering, drops attribute-value pairs from rules, and checks consistency “using Neighborhood Rough Set.” The paper states that the minimal number of attribute-value pairs are kept in the final rule. The resulting rules are applied to testing data to generate sentence importance, after which a summary of length $N:U\to\mathcal{P}(U)$3 words is produced. In some experiments, sentence ranking is further refined by the Aggregate-Rank-Measure from the cited prior work (Yadav, 2021).

The empirical findings are mixed but specific. For DUC2003 without Aggregate-Rank-Measure, classical LERS performs best among all methods, and NLER is competitive but not superior. For DUC2003 with Aggregate-Rank-Measure, $N:U\to\mathcal{P}(U)$4 performs best on ROUGE-1, ROUGE-L, and ROUGE-SU, while LERS performs best on ROUGE-2. For DUC2005, both without post-processing and with Aggregate-Rank-Measure, NLER ($N:U\to\mathcal{P}(U)$5 and $N:U\to\mathcal{P}(U)$6) is reported as performing best on all ROUGE metrics. The paper’s overall conclusion is that neighborhood rough set based techniques perform better than the normal rough-set based LERS algorithm, particularly when combined with Aggregate Rank Measure (Yadav, 2021).

These results are application-specific rather than theorem-driven. The same paper notes that its theoretical treatment of NRS is mainly definitional and does not provide separate theorems, formal robustness proofs, or a detailed complexity analysis (Yadav, 2021).

5. Unification with Pawlak rough sets through granular-ball models

A significant later development is the granular-ball rough set model, which is presented as a unified model of Pawlak rough set and neighborhood rough set. The stated objective is to retain NRS’s capacity to process continuous data while restoring equivalence-class-based knowledge representation (Xia et al., 2022).

In granular-ball computing, a granular-ball $N:U\to\mathcal{P}(U)$7 is a subset of samples

$N:U\to\mathcal{P}(U)$8

with center

$N:U\to\mathcal{P}(U)$9

and radius

$X\subseteq U$0

Granular-balls are generated by recursively splitting impure balls until purity conditions are met, and overlaps between heterogeneous balls are explicitly removed. A granular-ball can also be written as

$X\subseteq U$1

so each ball behaves as an adaptive neighborhood in metric space (Xia et al., 2022).

The model defines a granular-ball-based indiscernibility relation

$X\subseteq U$2

leading to equivalence classes

$X\subseteq U$3

Lower and upper approximations are then given in Pawlak form: $X\subseteq U$4

$X\subseteq U$5

For decision systems, positive and boundary regions are defined exactly as in classical rough sets, but the underlying granules are granular-balls rather than strict value-equivalence classes (Xia et al., 2022).

The paper’s unification claim is explicit. When the radius of each granular-ball is set to an infinitesimally small positive number, GBRS reduces to Pawlak rough sets; when the radius is not zero, GBRS becomes granular-ball NRS (GBNRS). This yields the following correspondence.

Model	Granulation basis	Stated role
Pawlak rough set	Equivalence classes	Discrete, equivalence-based knowledge
Neighborhood rough set	Metric neighborhoods	Continuous-data handling
Granular-ball rough set	Granular-ball equivalence classes	Unified representation of both

The experiments reported in the granular-ball paper use 6 discrete and 9 continuous datasets with 5-fold cross-validation and k-NN for evaluation. On the 6 discrete datasets, the average accuracies are reported as 0.8906 for the original data, 0.8618 for PRS, and 0.8958 for GBNRS. On the 9 continuous datasets, GBNRS is reported as achieving the highest classification accuracy on most datasets and outperforming classical NRS, weighted NRS, and several other feature-selection methods including Cfs, Ilfs, Laplacian Score, Lasso, and Mrmr (Xia et al., 2022).

The same paper also states a strong representational claim: because GBNRS uses adaptive granular-balls, avoids heterogeneous overlaps, and restores equivalence-class representation, it can serve as a representative method of neighborhood rough set (Xia et al., 2022).

6. Taxonomy, illustrative example, and terminological ambiguity

Within the broader rough-set landscape, neighborhood rough sets are treated as one of the main paradigms organized by granulation mechanism, alongside equivalence-based, tolerance-based, covering-based, and probabilistic models. The handbook places neighborhood models in settings where local granules derived from metrics, graphs, kNN relations, or similarity thresholds are more appropriate than global partitions. It also emphasizes applications such as predictive maintenance, local classification, anomaly screening, and dynamic or streaming contexts (Fujita et al., 7 Apr 2026).

A concrete handbook example uses predictive maintenance on vibration sensors. The universe is

$X\subseteq U$6

each sensor has a feature vector $X\subseteq U$7, and the distance is Euclidean: $X\subseteq U$8 With radius $X\subseteq U$9, the neighborhood system is defined by

$U$00

For the confirmed faulty-sensor set

$U$01

the handbook computes

$U$02

hence

$U$03

This example makes explicit the local semantics of NRS: $U$04 is definitely associated with the faulty region, $U$05 lie in the boundary, and $U$06 are definitely outside it (Fujita et al., 7 Apr 2026).

A recurrent misconception is to treat “rough neighborhood set” as if it always referred to a single formalism. The literature represented here shows instead that NRS is a family of models sharing the same lower/upper approximation schema but differing in how neighborhoods are generated: from coverings, metric balls, relations, similarities, or adaptive granular-balls (Yao et al., 2012, Fujita et al., 7 Apr 2026). A separate terminological ambiguity is that the abbreviation NRS is also used in an unrelated sense in the paper on NRS$U$07, where it denotes multidimensional Newton–Raphson–Simpson methods for polynomial root-finding. In that usage, NRS$U$08 concerns fixed points, Jacobians, and pairwise sums of polynomial zeros, not rough approximations or data granulation (DeFranco, 17 Sep 2025).

Taken together, these sources present Rough Neighborhood Set as a generalization of rough-set approximation from equivalence classes to local granules. Covering-based work clarifies when neighborhood systems recover partition behavior; metric formulations enable direct treatment of numeric data; application papers show how NRS can drive rule induction and supervised ranking; and granular-ball models attempt to unify neighborhood and Pawlak semantics within a single adaptive framework (Yao et al., 2012, Yadav, 2021, Xia et al., 2022, Fujita et al., 7 Apr 2026).