Introducing blockmodeling to input-output analysis

Michael Weber

EC3 – E-Commerce Competence Center

Donau-City-Strasse 1
A-1220 Vienna, Austria
michael.weber@ec3.at

Abstract. Complexity raises uncertainty and uncertainty calls for distinct actions to

reduce complexity and to establish an overview (i.e. order). This paper gives a review

on recent developments in a branch of research that aims at gaining insight into the

complex interweavements of relational data (graphs). It discusses a set of methods for

simultaneously clustering nodes and partitioning edges called blockmodeling.

Originally, blockmodeling was developed for applications in sociometry and

psychometry. Recent results of research, however, indicate the potentials of

blockmodeling for econometrics, input-output analysis in particular. Even though the

methods are in their infancy, it becomes apparent that blockmodeling provides an

interesting way to generate information to support the coordination of relations between

economic units. This might eventually benefit application fields of methods of input-

output analysis such as supply chain management. Extracting structural information

from relational data appears to become an important capability in an increasingly

‘networked’ economy and a precondition for supporting business collaboration.

Blockmodeling and especially a branch of blockmodeling research termed ‘generalized

blockmodeling’ might enhance the instruments of input-output analysts that are geared

to these – usually both exploratory and confirmatory – structure analytical purposes.

1. Introduction

Economic developments such as globalisation and the increasing differentiation of

processes for value generation that both are extensively supported by information

technologies lead to a progressive as well as world-spanning division of labour.

Economic units are subject to these developments and due to the division of labour also

reliant on coordination to collaborate and generate value. As the degree of division of

labour rises, so does the need of coordination for economic units and, consequently, the

set of (actual and potential) relations that are or could be established for the purpose of

collaboration (Weber & Fröschl, 2006). Evidently, this rising complexity as well as the

capabilities of present-day information systems to capture ever-growing extracts of

these high-dimensional data also influence the analysis of economic relations that

exhibit the economic transactions made. This development poses a considerable

challenge for the analysts that have to deal with these intricate and vast amounts of

relational data. Methods and procedures are needed that help to condense data to

retrieve relevant information. In the following, a contribution to this objective is

presented to input-output analysts that might also serve as a starting point for

considerations on the promotion of value generating linkages in supply chain

management. Applications could be conceived that assist economic units to mitigate

difficulties when navigating through the transaction system in an increasingly dynamic

economic environment by supporting business collaboration. Subsequently, a set of

methods for clustering relational data is discussed, starting with a general introduction

including definitions of types of equivalence relations and ideal blocks. Algorithms for

the generation of ‘optimal’ partitions of vertices and edges for binary and valued graphs

are then outlined and illustrated by simple examples.

2. Blockmodeling – clustering and partitioning of relational data

The identification and categorization of linkage patterns in relational data is an

insightful procedure that facilitates the analysis and interpretation particularly for graphs

(i.e. relational data) with a relatively high number of vertices, e.g. input-output data.

Building upon results of research in mathematical psychology and mathematical

sociology that was conducted primarily in the 1970s, recent research on methods for this

purpose in the field of social network analysis focuses on a specific approach that

simultaneously clusters the vertices of a graph and partitions the relations between the

vertices in so-called blocks. This approach – ‘blockmodeling’ (Breiger et al., 1975;

White et al., 1976) – evolved over the years starting from an indirect clustering

procedure that builds clusters based on structural (Lorrain & White, 1971; Burt, 1976)

and later even regular (White & Reitz, 1983) equivalence of vertices to an approach that

incorporates indirect and in particular also direct methods for building blockmodels

(Batagelj et al., 1992a; 1992b) with generalized concepts of equivalence (Doreian et al.,

1994; Batagelj, 1997).

Structural equivalence of two vertices v, w  V in a (directed) graph G(V,E) can be

identified in cases where every (directed) relation of vertex v can be matched to any

other vertex x  V by an (equally directed) relation of vertex w to vertex x. This means

that for every edge (v,x) and/or (x,v) a corresponding edge (w,x) and/or (x,w) exists with

v  w. In the blockmodel of a directed graph, the structurally equivalent vertices are

presented in a consolidated view. Regular equivalence is the generalization of structural

equivalence. In a (directed) graph G(V,E) two vertices v, w  V are regularly equivalent

if every (directed) relation of v to a vertex x  V is matched to an (equally directed)

relation of w to a vertex y  V. That is, for an edge (v,x) and/or (x,v), there also exists

the edge (w,y) and/or (y,w), with v  w. Both types of equivalence can be seen as a

starting point for the definition of groups of vertices. For instance, the vertices x and y

as well as the vertices v and w could be grouped and then the groups could be related to

each other.

2.1. Generating blockmodels

In recent years, a direct method for computing blockmodels was (further) elaborated

that is called ‘generalized blockmodeling’ (Doreian et al., 2004; Doreian et al., 2005;

building upon Batagelj et al., 1992a; 1992b; Doreian et al., 1994; Batagelj, 1997).

Particularly, Doreian et al. (2004) provide an interesting enhancement of this branch of

research. They introduce an extension of the (direct) blockmodeling approach that

enables the separate treatment and classification of the relations of a vertex of a graph to

its ancestor or successor vertices and therefore provide a way to cluster ancestor and

successor vertices as well as to partition the edges between these vertices differently.

This is an idea that was also discussed by Borgatti & Everett (1992). It enables

blockmodeling of graphs with disjoint sets of ancestor and successor vertices – so-called

‘two-mode data’ – as well as of graphs with one set of vertices (‘one-mode data’) with

each vertex having an ancestor and a successor role. A central challenge for direct,

‘two-mode’ blockmodels thereby is the necessity to ‘find structure’ of ancestor and

successor vertices with only a variable subset of the corresponding set of vertices

(successor or ancestor vertices) determining the group (cluster) membership. This is a

substantial difference to conventional clustering approaches that start from (two)

disjoint sets and their relations with one set encompassing the observation units

(vertices) while the other set contains the characteristics of the observation units. Using

measures of (dis-)similarity, these two-mode data are then transformed to a distance

matrix with the dimension of the number of observation units times the number of

observations units. This distance matrix can be seen as one-mode and forms the basis

for the clustering procedure chosen.

With this in mind, the distinction between indirect and direct blockmodeling can be

explained. Indirect blockmodeling approaches (e.g. Breiger et al., 1975) identify certain

attributes for each vertex – usually structural characteristics such as degree, reachability

of nodes or distance indices to successors or ancestors – that form the basis for the

generation of a distance matrix. This procedure corresponds to a transformation of the

original one-mode data to two-mode data (vertices and their characteristics) to

retransform the data back to a one-mode structure (a distance matrix). Based on the

distance matrix, conventional clustering methods are then employed. The direct

methods for blockmodeling (commencing with Batagelj et al., 1992a), however, avoid

this detour. Starting from a specification of an ideal blockmodel which incorporates

certain types of blocks, that vary according to the type of equivalence chosen, an

iterative optimization procedure is initiated to minimize the discrepancy of the

investigated empirical blockmodel, i.e. the blocks obtained by grouping the vertices,

and thus re-arranging the columns and rows of the empirical adjacency matrix, to the

ideal blockmodel. Two approaches can be distinguished when specifying an ideal

blockmodel: (i) an explorative (hypotheses-generating) approach that merely involves

the definition of the block types permitted as well as of the number of clusters or (ii) a

confirmatory approach that additionally implies the specific alignment of some or all

block types. This enables the testing of hypotheses on the structure of the relational

data. With both variations of direct blockmodeling the data is matched to an ideal

structure which stands in contrast to indirect blockmodeling that defines clusters

(groups of vertices) and blocks entirely ‘data-driven’, i.e. in a way that suits the data

best, and generates hypotheses about the block structure without any model pre-

specification.

2.2. Ideal block types

The definition of equivalence relations is an integral precondition for the specification

of an ideal blockmodel as it confines the block types admissible. These block types can

be understood as templates for the ideal blocks of the model. Block types for structural

equivalence are null blocks or complete blocks, while for regular equivalence the

standard block types are null blocks and regular blocks. However, in a generalized

blockmodel, additional block types can be defined (Doreian et al., 1994; Batagelj,

1997)). Examples and descriptions of such blocks are listed in Table 1.

<Insert: Table 1>

The relationship of the block types presented are visualised in Figure 1. It can be seen

that the definition of block types constitutes a certain hierarchy. Among the block types

for regular equivalence, the row-dominant and/or column-functional and the column-
dominant and/or the row-functional blocks permit the finest level of distinction.

Therefore, these four types build the most specific level of the block type hierarchy.

According to their definitions, the block types row-dominant and row-functional as well

as column-dominant and column-functional are mutually exclusive. The second tier of

the hierarchy in Figure 1 is formed by the block types column-regular and row-regular.

These types either appear independently from the most specific level as their definition

is more general, or they are derived from the corresponding types from the finest level

of the hierarchy. Thus, a row-dominant block or a column-functional block can also be

classified as column-regular. Similarly, a column-dominant block or a row-functional

block meets the requirements for row-regularity. However, a block that is classified as

row-dominant is not necessarily row-regular. Likewise, a column-dominant block

cannot automatically be categorized as column-regular. As can be told from Figure 1,

the most general block type in the hierarchy is the regular block. It can directly be

derived from the joint occurrence of column-regularity and row-regularity. Moreover, it

is possible to deduce it indirectly through a combination of types of blocks that can be

found in the third (finest) level of the hierarchy. For instance, the joint classification of a

block as column-functional and row-functional reveals that it is both, row-regular and

regular. Equally, a column-dominant and row-dominant block is not only column-

regular and row-regular, but also regular. In this context, the complete block, a special

case of the regular block that is used for the identification of structural equivalence, has

to be mentioned. Evidently, it is both, (outright) column- and row-dominant as well as

regular. A further special case of the basic block typology in Figure 1 is the null block,

which is in contrast to all the block types mentioned so far.

<Insert: Figure 1>

Building upon the knowledge of equivalence relations and the block types that can be
derived from these equivalences as well as the number of clusters desired, an ideal
blockmodel for (binary) relational data can be defined as follows (Table 2).

<Insert: Table 2>

3. Blockmodeling of binary data

3.1. An algorithm for direct one-mode blockmodeling

Direct blockmodeling basically quantifies and minimizes the divergence between an

empirical and an ideal blockmodel. To this end – subsequent to the designation of an

ideal blockmodel – an objective function has to be defined for the optimization

procedure as described in formula (1) (Batagelj et al., 1992a) for direct one-mode

blockmodeling.

(1)

The objective function Z(C) measures the discrepancy between the investigated

empirical blockmodel R(C) and the ideal model B(C), with defining the

set of disjoint clusters (groups of vertices) of the empirical blockmodel, i.e.

. The number of clusters (groups of vertices) k = |C| is

determined by the pre-specified ideal blockmodel. Initially, the ideal model is only
defined in terms of the number of clusters and the block types allowed. Then, through

the assignment of vertices to clusters, ideal blocks of the same dimension as the

empirical blocks are yielded that are formed with regard to the admissible block types.

In case of a confirmatory blockmodeling scenario, the objective function is defined as

the sum of discrepancies z(Ci,Cj) between each empirical block r(Ci,Cj)  R(C) and its

corresponding ideal block b, whereas in an exploratory scenario, the discrepancy, and

thus, the objective function, is based on the set of feasible ideal blocks B(Ci,Cj). In the

former scenario, the discrepancy z(Ci,Cj) of an empirical block r(Ci,Cj) to a

corresponding ideal block b is simply derived from the direct comparison of these two

blocks, i.e. it equals the measure of discrepancy (r(Ci,Cj),b). However, in the

exploratory scenario the discrepancy z(Ci,Cj) is calculated with reference to the set of

feasible ideal blocks B(Ci,Cj), i.e. z(Ci,Cj) is defined as the minimal discrepancy 

between the empirical block r(Ci,Cj) and all ideal blocks b  B(Ci,Cj) in line with the

admissible block types as specified in Formula (2).

(2)

Two examples for calculating the discrepancy  are given below. The suggestions

follow the ideas of Batagelj et al. (1992a) and (1992b).

(3)

Formula (3) features the designation of a blockmodel for structural equivalence. It

calculates the sum of the absolute deviations between the cell values rvw of the empirical

block and the corresponding cell values bvw in the ideal block. This implies equal

weights for positive and negative deviations. Alternatively, different weighting

coefficients can be assigned.

For regular equivalence, Formula (4) can be applied.

(4)

According to Formula (4), the discrepancy  is determined either as the sum of the cell

values rvw in the empirical block (if the ideal block b is null) or as the sum of (i) the

product of the number of zero columns and the total number of rows and (ii) the product

of the number of zero rows and the total number of columns (if the block type for b is

regular).
Building upon the definition of discrepancy, the optimization problem can be described

as follows. For a graph G(V,E), the vertices v  V shall be clustered in groups of

vertices Ci  C* under simultaneous partitioning of the relations e  E between the

vertices in a way that the objective function Z(C) is minimized (Formula (5)).

(5)

An iterative optimization procedure solves this problem by starting from an initial

(usually random) clustering C. This clustering is replaced by a clustering C´ that lies in

the neighbourhood of the current clustering C if the objective function value of C´ is

smaller than the corresponding value of the current clustering C. Doreian et al. (1994)

recommend two transformations in order to find new clusterings, either the shifting of a

vertex from one group of vertices to another or the exchange of two vertices between

two groups. This procedure is repeated until the objective function value reaches a

minimum. To avoid local optima, this routine is usually repeated severalfold with

varying initial clusterings.

3.2. An algorithm for direct two-mode blockmodeling

The designation of direct (generalized) blockmodels offers interesting opportunities to

discover structure and to validate hypotheses on a certain production system that can

condense in the relational data that reach beyond the analytical potential of methods
such as triangulation1 or simple graph theoretic ratios. This is especially true if the

blockmodel classifies vertices and partitions the relations by differentiating the role of

the vertices as ancestors and successors. For this reason, a direct blockmodeling

approach is required that can deal with two-mode data, i.e. data that relates different

vertices or vertices with differential interpretations2.

In a two-mode blockmodel, the definition of the objective function for the optimization

procedure changes as depicted in Formula (6) (Doreian et al., 2004).

(6)

Opposed to the one-mode approach, the ancestor and the successor vertices are grouped

in separate cluster sets. CZ contains the set of (empirical) groups of ancestor vertices,

whereas CS comprises the set of (empirical) groups of successor vertices. C is the tuple

of CZ and CS. The cardinalities of the groups of vertices k = |CZ| and h = |CS| are defined

by the pre-specified ideal blockmodel. For determining the discrepancy  between the

empirical blocks and the ideal blocks , the one-mode

formulas can be re-used by replacing Ci with and Cj with . Based on the

1 Triangulation is used for identifying the production hierarchy (production stages) by allowing for the transaction

volume of the (weighted) relational data in input-output analysis. It enables (quantitative) structural analysis, yet with

the drawback of an unstable order of production stages if the production relations are non-linear. This instability rises

with decreasing linearity of the relational data. It might also happen that (directly) unrelated vertices are strung

together through triangulation (Holub & Schnabl, 1985).

2 An example of such vertices is the distinction between the relations of a vertex to its ancestors and its successors.
discrepancy , the contributions to the objective function are calculated

according to formula (7).

(7)

The optimization problem for the two-mode approach can be described as follows. The

vertices of a graph G(V,E) that is made up of V = (V1,V2) and e = (v,w)  E with v  V1,

w  V2, are to be clustered to and , with , and

. Simultaneously the relations e  E of this graph are to be partitioned in a way

that minimizes the objective function Z(C) (Formula (8)).

(8)

Eventually, the two-mode approach applies the same iterative optimization procedure as

one-mode blockmodeling.

Assumptions on the ideal blockmodel require – at least partial – (prior) knowledge

about the structure of the relational data, i.e. the block types that are admissible as well

as the number of clusters. However, if one lacks this a priori information, approaches

have to be pursued that help to identify an ideal blockmodel and hence facilitate

exploratory blockmodeling. Brusco & Steinley (2006) present such an approach that
does not require any prior structural knowledge and that provides a basis for generating

a blockmodel for both, one- and two-mode relational data. Their approach builds upon a

column by column and row by row permutation of a quadratic or rectangular adjacency

matrix that aims at partitioning the relational data into homogeneous blocks. Thereby,

clusters are not directly identified, but can – for instance – be derived from the

visualisation of the results. The same is true for the determination of equivalence

relations or block types of the blocks that were formed. Unfortunately, the

computational performance of this approach that can be categorized as a branch and

bound method is dissatisfactory with graphs that have more than 40 vertices, as Brusco

& Steinley note by referring to alternative, heuristic methods that might improve the

performance of this approach in the future.

3.3. An illustrative example for direct two-mode blockmodeling

The generation of a two-mode blockmodel shall subsequently be demonstrated by a

simple example of the identification of structural equivalence. The formulas presented

above were therefore implemented in R (R, 2007) to enable the explorative designation

of a two-mode blockmodel by an iterative shifting of vertices. Starting point is a binary

adjacency matrix that consists of the row vertices V1 = {a, b, c, d, e, f} and the column

vertices V2 = {A, B, C, D, E} and relations as depicted in Table 3.

<Insert: Table 3>

As the structural equivalence of this adjacency matrix shall be examined, the algorithm

will optimize with reference to the (ideal) block types null and complete. The number of
row- and the number of column-clusters that is assumed for this demonstration is 3. (It

has to be noted that there could also be different numbers for row- and column-clusters

as the algorithm follows a two-mode approach.) Starting from an initial random

clustering as shown on the left-hand side of Figure 2, the algorithm produces a result

that perfectly corresponds to an ideal blockmodel for structural equivalence (right hand

side of Figure 2). The clusters and as well as and

form complete blocks and hence are structurally equivalent, while the

remaining combination of row- and column-clusters generates null blocks. The

objective function is 0, there is no discrepancy and the assumption of 3 row-clusters and

3 column-clusters is correct. Of course, this is just an example for the purpose of

demonstrating the approach. Larger and more complex data would make it more

difficult to find the ideal number of row- and column-clusters and would also involve a

higher number of optimization cycles to find a satisfactory solution.

<Insert: Figure 2>

4. Valued blockmodeling for input-output analysis

4.1. Block types and algorithms for valued blockmodeling

Evidently, domains that deal with valued relational data (weighted graphs) such as

input-output analysis might object to binary blockmodeling as the ‘binarization’ of data

leads to a considerable loss of information. Moreover, it seems to be difficult to define

an appropriate parameter (e.g. greater than 0 or the mean weight of the edges) for the
transformation of valued to binary relational data without influencing the solution

substantially. Nevertheless, blockmodeling that originates from applications dealing

with strictly binary data (Breiger et al., 1975) can facilitate the analysis of valued data

as well. This statement gains in importance in light of recent research contributions to

valued (generalized) blockmodeling that enable valued blockmodeling without

(significant) loss of information such as Žiberna (2007) or Weber & Denk (2007). By

building upon generalized blockmodeling (Doreian et al., 2005), Žiberna (2007)

develops two alternative approaches to blockmodel valued relational data. These are (i)

f-value and (ii) homogeneity blockmodeling. f-value blockmodeling (Žiberna (2007)

refers to it as ‘valued blockmodeling’) can be understood as a straightforward extension

of binary generalized one-mode blockmodeling that could readily be expanded to a two-

mode approach. It extends the equivalence relations and consequently the block type

definitions by replacing the stipulations for 1 with analogous stipulations for the

parameter m and by allowing for a function f of the cell values in a block that meets the

parameter m. Examples for these altered block type definitions can be found in Table 4;

the functions f that could be used for the valued blocks are, for instance, min(),

max(), sum(), or mean(). A further modification that is necessary to realize f-value

blockmodeling concerns the calculation of discrepancy between the empirical and the

ideal block types. Žiberna (2007) provides a distinct discrepancy criterion for each of

the altered block types. For a f-regular block, for instance, the discrepancy is derived

from the deviations of the function f evaluated for each column and each row to the

parameter m. More specifically, for each pair of a row vertex vk and a column vertex vl

the differences and are calculated. The

discrepancy of the block is then computed as sum of the pairwise maxima of dk. and d.l
over all pairs (vk,vl) for which the function f of the investigated row k of the investigated

column l is greater than the parameter m. In essence, these are the main adaptations of

the blockmodeling procedure for f-value blockmodeling. It has to be noted that –

compared to binary blockmodeling – a new issue has cropped up, viz. the determination

of the parameter m. It therefore becomes necessary to introduce additional procedures to

mitigate this problem, unless prior knowledge is available.

<Insert: Table 4>

As an alternative to f-value blockmodeling, Žiberna (2007) suggests to search for blocks

with a maximum homogeneity (‘homogeneity blockmodeling’) or minimum variability.

Therefore, a measure has to be established that records the homogeneity or variability

within a block such as the sum of squared deviations from the mean or the sum of

absolute deviations from the median. Moreover, the definition of the ideal block types

has to be adapted in a way that stresses the homogeneity (equality) of the main

characteristic of the block type in question (Table 4). This redefinition entails the

amendment of the discrepancy criterion for the block types as well. In comparison to

both, f-value blockmodeling and binary blockmodeling for valued data, no additional

parameters are needed for homogeneity blockmodeling and there is virtually no loss of

information. These are considerable advantages of this method over f-value

blockmodeling and binary blockmodeling of valued data. Still, this method is work in

progress as Žiberna (2007) notes. Thus, he suggests that the result of homogeneity

blockmodeling could be used as a starting solution for other approaches.

Weber & Denk (2007) propose a two-mode approach for valued blockmodeling. They

make use of block type definitions that allow for the mutual impact of vertices and

therefore capture the magnitude of the ‘inbound’ and ‘outbound’ effects of vertices

simultaneously. More emphasis is laid on block properties and cluster dependencies as

opposed to properties and dependencies of vertices as featured in conventional

blockmodeling. Consequently, their methodology is complementary to the ideas of

generalized blockmodeling that are summarized in Doreian et al. (2005). As for

Žiberna’s proposals, empirical studies that evaluate the expected benefits (and

shortcomings) of the valued blockmodeling approach suggested by Weber & Denk

(2007) are in an initial stage.

4.2. An illustrative example for direct two-mode blockmodeling of valued data

Figure 3 presents a demonstrative application of valued two-mode blockmodeling that

follows the homogeneity approach. From an input-output perspective, the (ancestor)

vertices can be interpreted as a set of commodities (labelled ‘a’, …, ‘j’), while the

(successor) vertices can be seen as a set of activities (labelled ‘A’, …,’H’). Evidently,

the edges display the valued relations between these two sets of vertices. Starting from

the valued adjacency matrix on the left-hand side homogeneous blocks are identified by

using the sum of absolute deviations from the block median as measure of within block

variability and max() as function f in the definition of f-regular blocks. In a max-

regular block the maximum cell value should be the same for each row and column.

Experiments with different numbers of row- and column-clusters revealed the suitability

of a blockmodel with 3 row-clusters and 4 column-clusters. The empirical blockmodel

resulting from the optimization procedure is displayed on the right-hand side of Figure

3. Besides four null-blocks containing unrelated commodities and activities, one

complete block was identified that is made up of three commodities with equal

importance for the block activity. The remaining blocks are max-regular with different

levels of cell values (relation weights) and discrepancies from the ideal max-regular

blocks.

<Insert: Figure 3>

It has to be emphasized that two-mode blockmodeling can analogously be applied to

(‘symmetric’) input-output tables (for instance tables with the dimensions commodities

x commodities), i.e. the vertex set V1 can be equal to the vertex set V2 (V1 = V2) or – to

put it differently – there is only one vertex set with two differing roles per vertex.

5. Conclusion

This paper provides an overview of ‘(generalized) blockmodeling’, a set of methods

designed to gain structural information from relational data, in order to facilitate the

following up of this approach for input-output analysts. Due to recent enhancements of

(generalized) blockmodeling, economic applications can be realized that might enrich

particularly input-output analysis as well as application fields such as supply chain

management. Blockmodeling could be furthered to an instrument that is able to analyse

and reduce the complexity of economic relations that rises with the increasing

availability of (transactional) data through information technology and the noticeable

‘networked’ economy. However, further research and improved algorithms are a

precondition to fully exploit the potentials of blockmodeling for econometric

applications. Particularly, large data sets exhaust the performance of present-day

computer systems and could therefore be a starting point for advanced algorithms, but

also the designation of blockmodels could be supported by specific routines and the

available set of block type definitions might be extended. Nevertheless, (generalized)

blockmodeling methods for the clustering and partitioning of relational data constitutes

an interesting branch of research that provides an attempt to deal with the complexity

inherent in data on (economic) linkages.

Acknowledgement

I would like to express my sincere gratitude to Michaela Denk for her comprehensive

support and our valuable discussions.

References

Batagelj, V. (1997) Notes on blockmodeling, Social Networks, 19(2), pp. 143-155.

Batagelj, V., Ferligoj, A., & Doreian, P. (1992a) Direct and indirect methods for

structural equivalence, Social Networks, 14(1-2), pp. 63-90.

Batagelj, V., Doreian, P., & Ferligoj, A. (1992b) An optimizational approach to regular

equivalence, Social Networks, 14(1-2), pp. 121-135.

Borgatti, S.P., & Everett, M.G. (1992) Regular blockmodels of multiway, multimode

matrices, Social Networks, 14(1-2), pp. 91-120.

Breiger, R.L., Boorman, S.A., & Arabie, P. (1975) An algorithm for clustering

relational data with applications to social network analysis, Journal of

Mathematical Psychology, 12, pp. 329–383.

Brusco, M., & Steinley, D. (2006) Inducing a blockmodel structure of two-mode binary

data using seriation procedures, Journal of Mathematical Psychology, 50, pp.

468–477.

Burt, R.S. (1976) Positions in networks, Social Forces, 55, pp. 93-122.

Doreian, P., Batagelj, V., & Ferligoj, A. (1994) Partitioning networks based on

generalized concepts of equivalence, The Journal of mathematical sociology,

19(1), pp. 1-27.

Doreian, P., Batagelj, V., & Ferligoj, A. (2004) Generalized blockmodeling of two-

mode network data, Social Networks, 26(1), pp. 29-53.

Doreian, P., Batagelj, V., & Ferligoj, A. (2005) Generalized Blockmodeling

(Cambridge University Press).

Holub, H.W., & Schnabl, H. (1985) Input-Output-Tabellen (München Wien,

Oldenbourg). In German.

Lorrain, F., & White, H.C. (1971) Structural equivalence of individuals in social

networks, Journal of Mathematical Sociology, 1, pp. 49-80.

R Development Core Team (2007) R: A language and environment for statistical

computing, R Foundation for Statistical Computing, Vienna, Austria.

Weber, M., & Denk, M. (2007) Two-mode blockmodeling for valued graphs, 31st

Annual Conference of the German Classification Society on Data Analysis,

Machine Learning, and Applications, March 7-9, 2007, Freiburg i. Br.,

Germany, Abstract booklet, p. 202.

Weber, M., & Fröschl, K. (2006) Das Semantic Web als Innovation in der

ökonomischen Koordination, in: T. Pellegrini & A. Blumauer (eds) Semantic

Web - Wege zur vernetzten Wissensgesellschaft (Berlin Heidelberg,

Springer), pp. 89-113. In German.

White, D.R., & Reitz, K.P. (1983) Graph and semigroup homomorphisms on networks

of relations, Social Networks, 5, pp. 193-234.

White, H.C., Boorman, S.A., & Breiger, R.L. (1976) Social structure from multiple

networks. I. Blockmodels of roles and positions, American Journal of

Sociology, 81, pp. 730-779.

Žiberna, A. (2007) Generalized blockmodeling of valued networks, Social Networks,

29(1), pp. 105-126.

 Table 1: Examples of block types for blockmodels based on Batagelj (1997)

adjacency matrix block type description

complete block
In a complete block, all vertices are directly connected.
That is, in binary adjacency matrix form, every cell of the
block holds the value 1. Self-reference of vertices can be
excluded from this as well as from the following
definitions.

regular block

In a regular block, each row and column of the binary

adjacency matrix has at least one cell with the value 1. In
other words, each vertex is related to at least one other
vertex.

null block

In a null block, there is no single pair of vertices that is

connected by an edge. Therefore, the cells of a null block
in a binary adjacency matrix are zero.

row-dominant block

A row-dominant block has the property that there is at

least one row in the block that features the value 1 in each
of its cells. Consequently, at least one ancestor is related to
every successor vertex that is assigned to the block.

row-regular block

A row-regular block holds at least one cell entry with the

value 1 in every row of the block. In other words, every
ancestor vertex is connected to at least one successor in the
block.

row-functional block

A row-functional block exhibits the value 1 exactly once in

each row. That is, every ancestor is exclusively connected
to one single successor. The row sums of the block in the
form of a binary adjacency matrix hold the value 1.
adjacency matrix block type description

column-dominant block

In a column-dominant block there is at least one column

that is completely covered with the value 1. In other
words, at least one successor is related to all (other)
ancestors in the block.

column-regular block

In a column-regular block, the sum of each column is at

least 1. Therefore, every successor is connected to at least
one ancestor in the block.

column-functional block

In a column-functional block, every column exhibits the

value 1 exactly once. That is, every successor is exclusively
related to one single ancestor in the block. The sum of
every column is exactly 1.
 Table 2: Example of an ideal blockmodel

C1 C2 C3 C4 C5
column- row-
C1 complete regular null
regular regular
row- column-
C2 null complete null
functional regular
row-
C3 regular null null null
functional
row- row- row-
C4 null null
regular regular functional
column- row- row-
C5 null null
regular regular functional
 Table 3: Sample data for two-mode blockmodeling

A B C D E
a 0 0 0 0 0
b 0 0 1 0 1
c 0 0 0 0 0
d 1 0 0 1 0
e 0 0 1 0 1
f 1 0 0 1 0
 Table 4: Examples for altered block types according to Žiberna (2007)

f-value blockmodeling homogeneity blockmodeling

In a f-regular block, the function f
returns the same value for each row
In a f-regular block, for each
as well as the same value for each
row and column of the valued
f-regular block column of the valued adjacency
adjacency matrix the function f
matrix. The results for the rows are
returns at least the value m.
not necessarily equal to the results
for the columns.
Every cell of the block from the
All cells within a complete block
complete block valued adjacency matrix holds
have the same value.
at least the value m.
In a row-f-regular block, the
The function f returns the same
row-f-regular function f returns at least the
value for every row in a row-f-
block value m in every row of the
regular block.
block.
A row-dominant block has at
In a row-dominant block there is at
row-dominant least one row in the block that
least one row within the block with
block exhibits at least the value m in
equal cell values.
each of its cells.
A row-functional block features All rows of a row-functional block
row-functional at least the value m exactly once have exactly one cell with the same
block in each row. The row sums of value that is greater than 0, all
the block are m. other cells are 0.
 null
complete
regular



column- row-
 regular  regular

row- 
dominant row-
column- column- functional
dominant functional




if ... then ... block types can emerge simultaneously

both ... and ...  joint occurrence of block type definitions

... can lead to ...  mutual exclusion of definitions

Figure 1: Relationship between block types

 D

D
E

B
a a

e c

c d

d f

b e

f b

Figure 2: Adjacency matrix before and after two-mode blockmodeling

 G

G
C

H
A

A
F

F
a 0 26 0 18 0 4 0 35 a 0 0 0 4 26 18 0 35

b 0 0 45 3 49 0 47 0 g 0 0 0 4 23 27 33 30

c 0 1 0 2 44 0 48 0 j 0 0 0 4 24 22 31 34

d 63 5 18 0 21 0 21 58 d 18 21 21 0 5 0 63 58

e 55 4 17 9 0 0 18 62 e 17 0 18 0 4 9 55 62

f 60 7 20 5 17 0 19 56 f 20 17 19 0 7 5 60 56

g 33 23 0 27 0 4 0 30 h 18 19 16 0 8 6 56 0

h 56 8 18 6 19 0 16 0 b 45 49 47 0 0 3 0 0

i 0 2 53 1 46 0 44 0 c 0 44 48 0 1 2 0 0

j 31 24 0 22 0 4 0 34 i 53 46 44 0 2 1 0 0

Figure 3: Valued adjacency matrix before and after valued two-mode blockmodeling