Elsevier

Information Sciences

Volume 501, October 2019, Pages 680-687
Information Sciences

On the degeneracy of the Randić entropy and related graph measures

https://doi.org/10.1016/j.ins.2018.11.011Get rights and content

Abstract

Numerous quantitative graph measures have been defined and applied in various disciplines. Such measures may be differentiated according to whether they are information-theoretic or non-information-theoretic. In this paper, we examine an important property of Randić entropy, an information-theoretic measure, and examine some related graph measures based on random roots. In particular, we investigate the degeneracy of these structural graph measures and discuss numerical results. Finally, we draw some conclusions about the measures’ applicability to deterministic and non-deterministic networks.

Introduction

Exploring the structure of complex networks has been a challenging problem over the last two decades [37], [42], [50]. Analysis of complex networks began, naturally enough, with descriptions of network properties. Random networks were distinguished from deterministic ones. Other considerations such as whether a network satisfies the “small world” hypothesis [50] or is scale free were also examined. Yet, these properties are relevant mainly to graphs representing the World Wide Web and various biological networks [9]. Another approach to analysis has been the quantification of structural properties based on measurement [15]. So-called graph measures can be divided into several categories such as global, local, information-theoretic, non-information-theoretic, distance-based and so forth [22], [28]. Finding a measure that performs well for a particular problem continues to be challenging [15]. Some progress has been made, however, as this approach has been pursued extensively in problem areas such as drug design, QSAR/QSPR, see, e.g., [3]. A critical aspect of this approach involves the definition of measures designed to characterize the structure of complex networks. Many different paths have been followed in quantitative graph analysis. Paths that has been followed for some time are the study of graph complexity [40] and comparative network analysis [23]. A more recent development involves network classification and clustering, see [46]. Critical to the use of quantitative methods to analyze networks are identifying the structural feature or features captured by a graph measure and determining the degeneracy of the measure [7], [19].

In this paper, we continue the work on the degeneracy of structural graph measures. Somewhat related to this problem is the search for complete graph invariants as attempted in [7], [16], [33], [47]. A complete set of graph invariants is a collection of properties which discriminate between all non-isomorphic graphs uniquely. In this paper, we focus on studying the degeneracy, i.e., the failure to discriminate non-isomorphic graphs uniquely utilizing information-theoretic and non-information-theoretic techniques. The work reported here builds on an earlier paper, namely, Dehmer et al. [20]. The point of departure here is Randić entropy [12], since the degeneracy of this measure has not yet been explored. To explore this measure we define a random polynomial using the so-called Randić weights [12], and exploit its roots as the basis for defining novel graph measures. The degeneracy of these graph measures is then investigated by computing numerical values for graphs and trees in an exhaustively set.

Section snippets

Information-theoretic graph measures

Note that a comprehensive review of the many extant graph measures (e.g., [6], [34], [48]) is beyond the scope of this paper. In any case, information-theoretic graph measures have been discussed extensively, see [40]. Here we just sketch two important classes of information-theoretic graph measures as they have been proven useful for network classification [22], [23] and network characterization [23], [40], [48].

A classical entropy based graph measure elaborated by Mowshowitz [40] built on the

Computational procedure

In this section, we present the details of a procedure designed to generate graph classes and to compute the novel graph measures.

  • Step 1:

    Classes of graphs Ti and Ni are generated exhaustively by means of Nauty [39]. Their sizes can be seen in Table 1. Ti is the class of all pairwise non-isomorphic trees with i vertices, and Ni is the class of all pairwise non-isomorphic graphs with i vertices. For computational details, see also [21]. The reason for using exhaustively generated trees and graphs is

Numerical results

The results of the computation of Randić entropy IR according to Eq. 9 are summarized in Table 1.

From Table 1 it is clear that the bigger the graph class, the higher the degeneracy rate. For example, 241842 of the graphs  ∈ N9 could not be distinguished structurally using the Randić entropy IR. In this case 99.92% (almost all) of the exhaustively generated graphs with 9 vertices cannot be discriminated by IR. The situation is quite similar for the classes of exhaustively generated trees. Even

Summary and conclusion

In this paper, we have examined the degeneracy of the Randić entropy and some related graph measures. Our starting point was the Randić entropy [12] which is based on the so-called Randić weights [12] for defining the probabilities on a given edge set. Also, we have defined additional graph measures based on the Randić polynomial as well as on the random Randić roots determined by Eqs. 12 and 14.

This approach is decidedly different from others dealing with ordinary graph measures used to

Acknowledgments

Matthias Dehmer thanks the Austrian Science Funds for supporting this work (project P 30031). Zengqiang Chen was supported by the National Natural Science Foundation of China (NSFC), see Grant No. 61573199. Yongtang Shi was partially supported by the National Natural Science Foundation of China (No. 11771221), the Natural Science Foundation of Tianjin (No. 17JCQNJC00300), the Natural Science Foundation of the Education Department of Shannxi ROC (No. 16JK1456 ). Chengyi Xia was supported by the

References (50)

  • D. Bonchev

    Information Theoretic Indices for Characterization of Chemical Structures

    (1983)
  • D. Bonchev

    Information theoretic measures of complexity

  • D. Bonchev et al.

    Isomer discrimination by topological information approach

    J. Comp. Chem.

    (1981)
  • M. Borgwardt

    Graph Kernels Ph.D. thesis

    (2007)
  • S. Bornholdt et al.

    Handbook of Graphs and Networks: From the Genome to the Internet

    (2003)
  • S.B. Bozkurt et al.

    Randić matrix and randić energy

    MATCH Commun. Math. Comp. Chem.

    (2010)
  • Z. Chen et al.

    Entropy of weighted graphs with randić weights

    Entropy

    (2015)
  • D.M. Cvetković et al.

    Spectra of Graphs. Theory and Application

    (1980)
  • M. Dehmer et al.

    Quantitative Graph Theory. Theory and Applications

    (2014)
  • M. Dehmer et al.

    Interrelations of graph distance measures based on topological indices

    PLoS ONE

    (2014)
  • M. Dehmer et al.

    The discrimination power of molecular identification numbers revisited

    MATCH Commun. Math. Comput. Chem.

    (2013)
  • M. Dehmer et al.

    An efficient heuristic approach to detecting graph isomorphism based on combinations of highly discriminating invariants

    Adv. Comput. Math.

    (2012)
  • M. Dehmer et al.

    Information indices with high discriminative power for graphs

    PLoS ONE

    (2012)
  • F. Emmert-Streib et al.

    Networks for systems biology: conceptual connection of data and function

    IET Syst. Biol.

    (2011)
  • F. Emmert-Streib et al.

    Classification of large graphs by a local tree decomposition

  • Cited by (0)

    View full text