Is-ClusterMPP: clustering algorithm through point processes and influence space towards high-dimensional data

Cluster Density Properties Define a Graph for Effective Pattern Feature Selection

“…A is the adjacency matrix corresponding to graph G, where each element a i,j represents a pairwise relationship between features F i and F j . Coefficients a i,j are defined via a potential function φ: heterogeneous densities [22]- [24]. This concept has been used as a new dimension, with the clustering process carried out in the new residual space.…”

“…Lv et al have used Is k to reduce the amount of DBSCAN input parameters [23]. Is k has also been used to determine the core-points, which supported the is-clustering algorithm in dense regions [24].…”

“…The density of neighbouring data around an observation x ∈ X can be estimated through the k-influence space (Is k (x)) [22]- [24]. The k-nearest neighbours set NN k (x) is never empty, whereas the size of RNN k (x) depends on how many times x is classified as k-nearest neighbor of an object x.…”

“…The UGFS scheme has shown a good performance, but it has two important limitations: (i) feature relevance to cluster discrimination is considered in the graph design, but not feature correlation, a redundancy indicator; and (ii) the k-nearest neighbors' set of each point is considered as a cluster representation, which can make this scheme sensitive to outliers and to the occurrence of clusters of unbalanced densities and shapes, which can lead to incorrect data description. The purpose of this study is to investigate a new unsupervised feature selection method, called Influence Space and Graph-based Feature Selection (ISGFS), which uses the k-influence space concept [22]- [24] and subspace learning to describe feature relationships and subsequently design a feature selection graph. The particularity of ISGFS is its ability to retrieve discriminative non-redundant features, applicable to data classes of arbitrary shape, size, and density.…”

“…Accordingly, ISGFS links features that preserve the local projected density in core data point neighborhoods. These neighborhoods, and cluster boundaries as well, can be approximated effectively via the k-influence space for arbitrary cluster layouts, including complex cluster shapes, unbalanced clusters, and clusters with significant overlap [22]- [24]. Moreover, graph edges are weighted according to feature correlation to account for feature redundancy in the selection process.…”