Predicting Graph Operator Output

over Multiple Graphs

Tasos Bakogiannis1(B) , Ioannis Giannakopoulos1 , Dimitrios Tsoumakos2 ,

and Nectarios Koziris1
1
CSLab, School of ECE, National Technical University of Athens, Athens, Greece
{abk,ggian,nkoziris}@cslab.ece.ntua.gr
2
Department of Informatics, Ionian University, Corfu, Greece
[email protected]

Abstract. A growing list of domains, in the forefront of which are Web

data and applications, are modeled by graph representations. In content-
driven graph analytics, knowledge must be extracted from large numbers
of available data graphs. As the number of datasets (a different type of
volume) can reach immense sizes, a thorough evaluation of each input
is prohibitively expensive. To date, there exists no efficient method to
quantify the impact of numerous available datasets over different graph
analytics tasks. To address this challenge, we propose an efficient graph
operator modeling methodology. Our novel, operator-agnostic approach
focuses on the inputs themselves, utilizing graph similarity to infer knowl-
edge about them. An operator is executed for a small subset of the avail-
able inputs and its behavior is modeled for the rest of the graphs utiliz-
ing machine learning. We propose a family of similarity measures based
on the degree distribution that prove capable of producing high quality
models for many popular graph tasks, even compared to modern, state
of the art similarity functions. Our evaluation over both real-world and
synthetic graph datasets indicates that our method achieves extremely
accurate modeling of many commonly encountered operators, managing
massive speedups over a brute-force alternative.

Keywords: Graph analytics · Operator modeling · Graph similarity

1 Introduction
A huge amount of data originating from the Web can be naturally expressed
as graphs, e.g., product co-purchasing graphs [25], community graphs [28], etc.
Graph analytics is a common tool used to effectively tackle complex tasks such as
social community analysis, recommendations, fraud detection, etc. Many diverse
graph operators are available [12], with functionality including the computation
of centrality measures, clustering metrics or network statistics [8], all regularly
utilized in tasks such as classification, community detection and link prediction.
Yet, as Big Data technologies mature and evolve, emphasis is placed on areas
not solely related to data (i.e., graph) size. A different type of challenge steadily
c Springer Nature Switzerland AG 2019
M. Bakaev et al. (Eds.): ICWE 2019, LNCS 11496, pp. 107–122, 2019.
https://doi.org/10.1007/978-3-030-19274-7_9
108 T. Bakogiannis et al.

shifts attention to the actual content. In content-based analytics [14], data is

processed for sense-making. Similarly, in content-sensitive applications the qual-
ity of insights derived is mainly attributed to the input content. The plethora of
available sources for content-sensitive analytics tasks now creates an issue: Data
scientists have to decide which of the available datasets will be used for a given
workflow, in order to maximize its impact. Yet, as modern analytics tasks have
evolved into increasingly long and complex series of diverse operators, evaluat-
ing the utility of immense numbers of inputs is prohibitively expensive. This is
notably true for graph operators, whose computational cost has led to extensive
research on approximation algorithms (e.g., [11,30]).
As a motivating example, let us consider a dataset consisting of a very large
number of citation graphs. We wish to identify those graphs that have the most
well-connected citations and contain highly-cited papers. As a result, the clus-
tering coefficient [8], a good measure of neighborhood connectivity, would have
to be computed for all the graphs in the dataset in order to allow the identifi-
cation of the top-k such graphs. To quantify the importance of each paper, we
consider a centrality measure such as betweenness centrality [8]. Consequently,
we would have to compute the maximum betweenness centrality score for each
citation graph and combine the results with those obtained from the analysis
based on the clustering coefficient. Yet, this could be a daunting task due to the
operators’ complexity and the size of the dataset.
The challenge this work tackles is thus the following: Given a graph analytics
operator and a large number of input graphs, can we reliably predict operator
output at low cost? In this work, we introduce a novel, operator-agnostic dataset
profiling mechanism. Rather than executing the operator over each input graph,
our work leverages the relationship between the dataset’s graphs, expressed
through a similarity measure, and infers knowledge about them. In our exam-
ple, instead of exhaustively computing the clustering coefficient, we calculate a
similarity matrix for our dataset, compute the clustering coefficient for a small
subset of graphs and utilize the similarity matrix to estimate its value for the
remaining graphs. We may then compute the maximum betweenness central-
ity for also a small subset of citation graphs and reuse the already calculated
similarity matrix to estimate the scores for the rest of the graphs.
Our method is based on the intuition that, for a given graph operator, sim-
ilar graphs produce similar outputs. This intuition is solidly supported by the
existence of strong correlations between different graph operators ([7,20,22]).
Hence, by assuming a similarity measure that correlates to a set of operators,
we can use machine learning techniques to approximate their outcomes. Given a
graph dataset and an operator to model, our method utilizes a similarity mea-
sure to compute the similarity matrix of the dataset, i.e., all-pairs similarity
scores between the graphs of the dataset. The given operator is then run for a
small subset of the dataset; using the similarity matrix and the available opera-
tor outputs, we are able to approximate the operator for the remaining graphs.
 Predicting Graph Operator Output over Multiple Graphs 109

To the best of our knowledge, this is the first effort to predict graph operator
output over large datasets. In summary, we make the following contributions:

– We propose a novel, similarity-based method to estimate graph operator

output for large graph datasets. This method shifts the complexity of numer-
ous graph computations to less expensive graph similarities. This choice offers
two major advantages: First, our scheme is operator-agnostic since the com-
puted similarity matrix can be reused. As a result, the similarity matrix com-
putation is amortized and the cost of our method is ultimately dominated by
the computation of that operator for a small subset of the dataset. Second, the
method is agnostic to the similarity measure that is used. This property gives
us the ability to utilize or arbitrarily combine different similarity measures.
– We introduce a family of similarity measures based on degree distribution with
a gradual trade-off between detail and computational complexity. Despite
their simplicity, they prove capable of producing highly accurate models,
comparable or even surpassing other more costly, state-of-the-art similarity
measures ([32,35]).
– We offer an open-source implementation1 of our method and perform an
extensive experimental evaluation using both synthetic and real datasets.
Our results indicate that we can accurately model a variety of popular graph
operators, with errors that can be < 1%, sampling a mere 5% of the graphs
for execution. Amortizing the similarity cost over six operators, the process
can produce up to 18× speed-up. Our proposed similarity measures produce
comparable or more accurate results to state-of-the-art similarity measures
but run more than 5 orders of magnitude faster.

2 Methodology

In this section, we formulate the problem and describe the methodology along
with different aspects of the proposed solution. We start off with some basic
notation followed throughout the paper and a formal description of our method
and its complexity.
Let a graph G be an ordered pair G = (V, E) with V being the set of vertices
and E the set of edges of G, respectively. The degree of a vertex u ∈ V , denoted
by dG (u), is the number of edges of G incident to u. The degree distribution of a
graph G, denoted by PG (k), expresses the probability that a randomly selected
vertex of G has degree k. A dataset D is a set of N simple, undirected graphs
D = {G1 , G2 , ..., GN }. We define a graph operator to be a function g : D → R,
mapping an element of D to a real number. In order to quantify the similarity
between two graphs Ga , Gb ∈ D we use a graph similarity function s : D×D → R
with range within [0, 1]. For two graphs Ga , Gb ∈ D, a similarity of 1 implies
that they are identical while a similarity of 0 the opposite.
Consequently, the problem we are addressing can be formally stated as fol-
lows: Given a dataset of graphs D and a graph operator g, without knowledge of
the range of g given D, we wish to infer a function ĝ : D → R that approximates
1
https://github.com/giagiannis/data-profiler.
110 T. Bakogiannis et al.

g. Additionally, we wish our approximation to be both accurate (i.e., |g − ĝ| < ,

for some small ) and efficient (i.e., O(ĝ) < O(g)). In this formulation, our goal
is to provide an accurate approximation of g, while avoiding its exhaustive exe-
cution over the entire D. To achieve this goal, we utilize the similarity matrix R,
an N × N matrix with R[i, j] = s(Gi , Gj ), where s is a given similarity measure.
As a result, R contains all-pairs similarity scores between the graphs of D. R is
symmetric, its elements are in [0, 1] and the entries of its main diagonal equal 1.
Our method takes as input a dataset D and an operator g to model. It forms
a pipeline that begins with the computation of the similarity matrix R based
on s; calculates the actual values of g for a ratio p ∈ (0, 1) of randomly selected
graphs of D and, finally, estimates g for the remaining graphs of D by running a
weighted version of the k-Nearest-Neighbors (kNN) algorithm [19]. The inferred
function ĝ is then given by the following equation:

i∈Γk (x) wxi g(Gi )
ĝ(Gx ) =  (1)
i∈Γk (x) wxi

Where wxi = R[x, i] is the similarity score for graphs Gx , Gi , i.e., wxi =
s(Gx , Gi ), Γk (x) is the set of the k most similar graphs to Gx for which we
have already calculated g and g(Gi ) the value of the operator for Gi . Our app-
roach is formally described in Algorithm 1. The complexity of Algorithm 1 can
be broken down to its three main components: (1) The calculation of the simi-
larity matrix R in lines 3−4, for a given similarity measure s with complexity S.
(2) The component which computes the operator g for pN graphs (lines 5−7),
assuming that g has complexity M . And 3) the approximation of the operator
for the remaining graphs (lines 8−10) using kNN. Thus, the overall complexity
of our method is:

O(N 2 S + pN M + (N (1 − p))((pN )log(pN ) + k)) (2)

From Eq. 2, we deduce that the complexity of our method is dominated by its
first two components. Consequently, the lower the computational cost of s, the
more efficient our approach will be. Additionally, we expect our training set to
be much smaller than our original dataset (i.e., p  1).

Algorithm 1. Graph Operator Modeling

1: procedure Approximate([G1 , G2 , ..., GN ], g, s, p, k)
2: R ← [ ], T ← { }, A ← { }
3: for (i, j) ← [1, N ] × [1, N ] do
4: R[i, j] ← s(Gi , Gj )
5: for i ← 1, p · N do
6: r ← randint(1, N )
7: T [Gr ] ← g(Gr )
8: for x ← [G1 , G2 , ..., GN ], x ∈
/ keys(T ) do
9: t ← f indN eighbors(R, T, k, x)
10: A[x] ← calcApproximation(R, t)
11: return A
 Predicting Graph Operator Output over Multiple Graphs 111

It is important to note here that the O(N 2 S) component corresponds to a cal-

culation performed only once, whether modeling a single or multiple operators.
Thus, given that the similarity matrix calculation happens once per dataset, its
cost gets amortized over multiple graph operators, making the O(pN M ) factor
the dominant one for our pipeline.

2.1 Similarity Measures

The similarity matrix is an essential tool in our effort to model graph opera-
tors under the hypothesis that similar graphs produce similar operator outputs.
Relative to graph analytics operators, we propose a family of similarity mea-
sures based on graph degree distribution. Reinforced by the proven correlations
between many diverse graph operators ([7,20,22]), we intend the proposed sim-
ilarity measures to express graph similarity in a way that enables modeling of
various operators.
 Degree Distribution: In order to quantify the similarity between two graphs
we rely on comparing their degree distributions. We choose the Bhattacharyya
coefficient BC [5] to perform this task. BC is considered a highly advantageous
method for comparing distributions [1]. BC divides the output space of a function
into m partitions and uses the cardinality of each partition to create an m-
dimensional vector representing that space. As a measure of divergence between
two distributions, the square of the angle between the two vectors is considered.
In our case, the points of the output space are the degrees of the nodes of each
graph. By dividing that space into partitions and considering the cardinalities of
each partition, we effectively compare the degree distributions of those graphs.
In our implementation, we use a k-d tree [4], a data structure used for space
partitioning, to compute BC. We build a k-d tree once, based on a predefined
percentage of vertex degrees from all the graphs in D. We then use the created
space partitioning to compute degree distributions for each graph.
 Degree Distribution + Levels: As an extension of the degree distribution-
based similarity measure, we consider a class of measures with increasing level
of information. Intuitively, the degree of a vertex is a measure of its connectivity
based on its immediate neighbors. Adding a level of indirection, we can consider
the degree of a vertex at level 1 as the degree of a super-node containing the
vertex and all its immediate neighbors. Generalizing this idea to more than one
levels gives us a measure of the indirect connectivity of a vertex. By combin-
ing the degrees of a vertex for multiple levels we obtain information up to level
hops away. As an illustrative example, in Fig. 1 vertex u0 has degree 4, when
considering its direct neighbors, 1 when its neighborhood is expanded to level
1, and for level 2 it becomes 3. As a result, we quantify the similarity between
graphs by calculating the degrees up to a certain level and use BC to compare
the resulting degree distributions. A good property of this class of measures is
that they provide us with a nice trade-off between accuracy and computational
cost. Increasing the number of degree distribution levels involves additional com-
putations but also incorporates more graph topological insights to it. In order to
112 T. Bakogiannis et al.

calculate the degrees for a given level, for each vertex we perform a depth-limited
Depth First Search up to level hops away in order to mark the internal edges
of the super-node. We then count the edges of the border vertices (vertices level
hops away from the source) that do not connect to any internal vertices.
 Degree Distribution + Vertex Count: A second extension to our degree
distribution-based similarity measure is based on the ability of our method to
combine similarity matrices. Graph size (vertex count) is another graph attribute
to measure similarity on. We formulate similarity in terms of vertex count
min(|VG |,|VG |)
as: s(Gi , Gj ) = max(|VGi |,|VGj |) . Intuitively, s approaches 1 when |VGi | − |VGj |
i j
approaches 0, i.e., when Gi , Gj have similar vertex counts. To incorporate vertex
count into the graph comparison, we can combine the similarity matrices com-
puted with degree distributions and vertex counts using an arbitrary formula
(e.g., linear composition).

u8
u7
u0
u5 u1
=4
Level 0
Leve
Level 2 = 3 l1=
1
u4 u2
u6 u3

Fig. 1. Example of Degree Distribution + Levels

2.2 Discussion

In this section, we consider a series of issues that relate to the configuration and
performance of our method as well as to the relation between modeled operators,
similarity measure and input datasets.
 Graph Operators: This work focuses on graph analytics operators, namely
centralities, clustering metrics, network statistics, etc. Research on this area has
resulted in a large collection of operators, also referred to as topology metrics
(e.g., [7,8,12,20]). Topology metrics can be loosely classified in three categories
([7,20,22]), those related to distance, connectivity and spectrum. In the first
class, we find metrics like diameter, average distance or betweenness centrality.
In the second, average degree, degree distribution, etc. Finally, the third class
comes from the spectral analysis of a graph and contains the computation of
eigenvalues, eigenvectors or other spectral-related metrics.
 Combining Similarity Measures: We can think of use cases where we
want to quantify the similarity of graphs based on parameters unrelated to each
other. For example, we might want to compare two graphs based on their degree
distributions but also take under account their vertex count. This composition
can be naturally implemented in our system by computing independent similarity
 Predicting Graph Operator Output over Multiple Graphs 113

matrices and “fuse” those matrices into one using a formula. This technique is
presented in our evaluation and proves effective for a number of operators.
 Regression Analysis: Although there exist several approaches to statistical
learning [19], we have opted for the kNN method. We choose kNN for its sim-
plicity and because we do not have to calculate distances between points of our
dataset (we already have that information from the similarity matrix). The kNN
algorithm is also suitable for our use case since it is sensitive to localized data and
insensitive to outliers. A desired property, since we expect similar graphs to have
similar operator scores and should therefore be of influence in our estimations.
 Scaling Similarity Computations: Having to compute all-pairs similarity
scores for a large collection of graphs can be prohibitively expensive. To this end,
we introduce a preprocessing step which we argue that improves on the existing
computational cost, reducing the number of similarity calculations performed.
As, in order to approximate a graph operator, we employ kNN, we observe
that, for each graph, we only require the similarity scores to its k most similar
graphs for which we have the value of g, i.e., the weights in Eq. 1. Therefore we
propose to run a clustering algorithm which will produce clusters of graphs with
high similarity. Then for each cluster compute all-pairs similarity scores between
its members, setting inter-cluster similarities to zero. By creating clusters of
size much larger than k, we expect minimal loss in accuracy while avoiding a
considerable number of similarity computations. As a clustering algorithm we
use a simplified version of k-medoids in combination with k-means++, for the
initial seed selection ([2,23]). For an extensive experimental evaluation of this
technique we refer the reader to the extended version of our work in [34] which
we have not included here due to space constraints.

3 Experimental Evaluation

 Datasets: For our experimental evaluation, we consider both real and syn-
thetic datasets. The real datasets comprise a set of ego graphs from Twitter
(TW ) which consists of 973 user “circles” as well as a dataset containing 733
snapshots of the graph that is formed by considering the Autonomous Systems
(AS ) that comprise the Internet as nodes and adding links between those sys-
tems that communicate to each other. Both datasets are taken from the Stanford
Large Network Dataset Collection [26].
We also experiment with a dataset of synthetic graphs (referred to as the
BA dataset) generated using the SNAP library [27]. We use the GenPrefAttach
generator to create random scale-free graphs with power-law degree distributions
using the Barabasi-Albert model [3]. We keep the vertex count of the graphs
constant to 4K. We introduce randomness to this dataset by having the initial
outdegree of each vertex be a uniformly random number in the range [1, 32]. The
Barabasi-Albert model constructs a graph by adding one vertex at a time. The
initial outdegree of a vertex is the maximum number of vertices it connects to,
114 T. Bakogiannis et al.

the moment it is added to the graph. The graphs of the dataset are simple and
undirected. Further details about the datasets can be found in Table 1.
 Similarity Measures: We evaluate all the similarity measures proposed in
Sect. 2.1, namely degree distribution + levels, for levels 0, 1, 2 and degree dis-
tribution + vertex count. When combining vertex count with degree, we use
the following simple formula: R = w1 Rd + w2 Rn , with Rd , Rn the degree dis-
tribution and vertex count similarity matrices respectively. In our evaluation,
w1 = w2 = 0.5. To investigate their strengths and limitations, we compare them
against two measures functioning as our baselines. The first is a sophisticated
similarity measure not based on degree but rather on distance distributions (from
which the degree distribution can be deduced). D-measure [32] is based on the
concept of network node dispersion (NND) which is a measure of the heterogene-
ity of a graph in terms of connectivity distances. It is a state-of-the-art graph
similarity measure with very good experimental results for both real and syn-
thetic graphs. Our second baseline comes from the extensively researched area
of graph kernels. For the purposes of our evaluation, we opted for the geometric
Random Walk Kernel (rw-kernel ) [16] as a widely used representative of this
class of similarity measures. In order to avoid the halting phenomenon due to
the kernel’s decay factor (λk ) we set λ = 0.1 and the number of steps k ≤ 4,
values that are considered to be reasonable for the general case [33].

Table 1. Datasets overview

Name Size (N) |V | |E| Range |V | Range |E|

TW 973 132 1,841 min: 6 min: 9
max: 248 max: 12,387
AS 733 4,183 8,540 min: 103 min: 248
max: 6,474 max: 13,895
BA 1,000 4,000 66,865 4,000 min: 3,999
max: 127,472

 Graph Operators: In our evaluation, we model operators from all the

categories mentioned in Sect. 2.2. As representatives of the distance class, we
choose betweenness (bc), edge betweenness (ebc) and closeness centralities (cc)
([8,29]), three metrics that express how central a vertex or edge is in a graph.
From the spectrum class, we choose spectral radius (sr) and eigenvector cen-
trality (ec). The first is associated with the robustness of a network against the
spreading of a virus [21], while the second also expresses vertex centrality [6].
Finally, as a connectivity related metric we consider PageRank (pr), a centrality
measure used for ranking web pages based on popularity [9].
All measures, except spectral radius, are centrality measures expressed at
vertex level (edge level in the case of edge betweenness). Since we wish all our
 Predicting Graph Operator Output over Multiple Graphs 115

measures to be expressed at graph level, we will be using a method attributed

to Freeman [13] to make that generalization. This is a general approach that
can be applied to any centrality [8], and measures the average difference in
centrality between the most central point and all others. All the graph operators
are implemented in R. We use the R package of the igraph library [10] which
contains implementations of all the algorithms mentioned.
 kNN: The only parameter we will have to specify for kNN is k. After extensive
experimentation (omitted due to space constraints), we have observed that small
values of k tend to perform better. As a result, all our experiments are performed
with k = 3.
 Error Metrics: The modeling accuracy of our method is quantified using two
widely used measures from the literature, the Median Absolute Percentage Error
and the Normalized Root Mean Squared Error.
 Setup: All experiments are conducted on an Openstack VM with 16 Intel Xeon
E312 processors at 2 GHz, 32 GB main memory running Ubuntu Server 16.04.3
LTS with Linux kernel 4.4.0. We implemented our prototype in Go language
(v.1.7.6).

3.1 Experiments

 Modeling Accuracy: To evaluate the accuracy of our approximations, we

calculate MdAPE and nRMSE for a randomized 20% of our dataset. We vary
the sampling ratio p, i.e., the number of graphs for which we actually execute
the operator, divided by the total number of graphs in the dataset. The results
are displayed in Table 2. Each row represents a combination of a dataset and
a graph operator with the corresponding error values for different values of p
between 5% and 20%.
The results in Table 2 showcase that our method is capable of modeling dif-
ferent classes of graph operators with very good accuracy. Although our app-
roach employs a degree distribution-based similarity measure, we observe that
the generated similarity matrix is expressive enough to allow the accurate mod-
eling of distance- and spectrum-related metrics as well, achieving errors well
below 10% for most cases. In AS graphs, the M dAP E error is less than 3.2%
for all the considered operators when only a mere 5% of the available graphs
is examined. Operators such as closeness or eigenvector centralities display low
M dAP E errors in the range of < 8% for all datasets. Through the use of more
expressive or combined similarity measures, our method can improve on these
results, as we show later in this Section. We also note that the approximation
accuracy increases with the sampling ratio. This is expressed by the decrease of
both M dAP E and nRM SE when we increase the size of our training set. These
results verify that modeling such graph operators is not only possible, but it can
also produce highly accurate models with marginal errors.
Specifically, in the case of the AS dataset, we observe that all the opera-
tors are modeled more accurately than in any other real or synthetic dataset.
116 T. Bakogiannis et al.

Table 2. Modeling errors and execution speedup

MdAPE (%) nRMSE Speedup × A. Speedup ×

5% 10% 20% 5% 10% 20% 5% 10% 20% 5% 10% 20%
AS sr 1.3 1.1 0.9 0.05 0.03 0.02 6.4 3.8 3.3 18.0 9.5 4.9
ec 0.1 0.1 0.0 0.01 0.00 0.00 5.7 4.5 3.1
bc 1.4 1.2 1.1 0.04 0.03 0.03 15.7 8.8 4.7
ebc 3.1 2.7 2.4 0.04 0.04 0.04 17.3 9.3 4.8
cc 0.4 0.4 0.3 0.01 0.01 0.01 14.0 8.2 4.5
pr 0.9 0.8 0.7 0.05 0.04 0.03 5.7 4.4 3.1
TW sr 16.3 15.3 14.7 0.10 0.10 0.10 13.3 8.0 4.4 14.8 8.5 4.6
ec 8.0 7.7 7.7 0.14 0.14 0.13 13.1 7.9 4.4
bc 17.8 17.5 16.8 0.16 0.15 0.14 13.0 7.8 4.4
ebc 29.5 29.8 28.6 0.12 0.12 0.12 13.5 8.0 4.4
cc 3.3 3.0 2.9 0.10 0.10 0.09 13.0 7.9 4.4
pr 9.2 7.7 7.2 0.07 0.06 0.05 13.2 7.9 4.4
BA sr 3.3 1.8 0.9 0.04 0.03 0.03 5.6 4.4 3.0 16.3 9.0 4.7
ec 0.4 0.3 0.3 0.01 0.01 0.01 3.7 3.1 2.4
bc 10.3 10.1 9.6 0.10 0.05 0.02 12.6 7.7 4.4
ebc 10.9 9.3 8.5 0.10 0.09 0.01 13.6 8.1 4.5
cc 2.4 2.2 2.1 0.04 0.04 0.03 9.9 6.6 4.0
pr 6.7 6.1 5.9 0.06 0.05 0.05 3.6 3.0 2.3

This can be attributed to the topology of the AS graphs. These graphs display
a linear relationship between vertex and edge counts. Their clustering coefficient
displays very little variance, suggesting that as the graphs grow in size they
keep the same topological structure. This gradual, uniform evolution of the AS
graphs leads to easier modeling of the values of a given graph topology measure.
On the other hand, our approach has better accuracy for degree- than
distance-related metrics in the cases of the TW and BA datasets. The simi-
larity measure we use is based on the degree distribution that is only indirectly
related to vertex distances. This can be seen, for example, in the case of BA if we
compare the modeling error for the betweenness centrality (bc) and PageRank
(pr) measures. Overall, we see that eigenvector and closeness centralities are the
two most accurately approximated metrics across all datasets. Next up, we find
PageRank, spectral radius, betweenness and edge betweenness centralities. Will-
ing to further examine the connection between modeling accuracy and similarity
measures, we have included D-measure and rw-kernel in our evaluation as well
as the degree-level similarity measures and the similarity matrice combination
technique.
 Execution Speedup: Next, we evaluate the gains our method can provide
in execution time. The similarity matrix computation is a time-consuming step,
 Predicting Graph Operator Output over Multiple Graphs 117

yet an advantage of our scheme is that the matrix can be reused for different
graph operators and thus its cost can be amortized. In order to provide a better
insight, we calculate two types of speedups: One that considers the similarity
matrix construction from scratch for each operator separately (provided in the
Speedup column of Table 2) and one that expresses the average speedup for all
six measures for each dataset, where the similarity matrix has been constructed
once (provided in the A. Speedup column of Table 2).
The observed results highlight that our method is not only capable of pro-
viding models of high quality, but also does so in a time-efficient manner. A
closer examination of the Speedup columns shows that our method is particu-
larly efficient for complex metrics that require more computation time (as in the
ebc and cc cases for all datasets). The upper bound of the theoretically antici-
pated speedup equals p1 , p being the sampling ratio. Interestingly, the Amortized
Speedup column indicates that when the procedure of constructing the similar-
ity matrix is amortized to the six operators under consideration, the achieved
speedup is very close to the theoretical one. This is indeed the case for the AS
and BA datasets that comprise the largest graphs, in terms of number of ver-
tices: For all p values, the amortized speedup closely approximates p1 . In the case
of the TW dataset which consists of much smaller graphs and, hence, the time
dedicated to the similarity matrix estimation is relatively larger than the previ-
ous cases, we observe that the achieved speedup is also sizable. In any case, the
capability of reusing the similarity matrix, which is calculated on a per-dataset
rather than on a per-operator basis, enables our approach to scale and be more
efficient as the number and complexity of graph operators increases.

0.6 0.6 0.6 0.6

0.5 0.5 0.5 0.5

0.4 0.4 0.4 0.4

MdAPE

MdAPE

MdAPE

MdAPE

0.3 0.3 0.3 0.3

0.2 0.2 0.2 0.2

0.1 0.1 0.1 0.1

0 0 0 0
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.05 0.1 0.15 0.2 0.25 0.3 0.35
Sampling Ratio Sampling Ratio Sampling Ratio Sampling Ratio
level-0 level-2 d-sim level-0 level-2 d-sim level-0 level-2 d-sim level-0 level-2 d-sim
level-1 level-0+size rw-kernel level-1 level-0+size rw-kernel level-1 level-0+size rw-kernel level-1 level-0+size rw-kernel

(a) Spectral Rad. (b) Eigenvector C. (c) Betweenness C. (d) Edge B. C.

0.6 0.6
6 level-0
Execution Time (sec)

0.5 0.5 10 level-1

0.4 0.4 level-2
MdAPE

MdAPE

4 d-sim
0.3 0.3 10 rw-kernel
0.2 0.2

0.1 0.1 102

0 0
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.05 0.1 0.15 0.2 0.25 0.3 0.35 1
Sampling Ratio Sampling Ratio
level-0 level-2 d-sim level-0 level-2 d-sim
level-1 level-0+size rw-kernel level-1 level-0+size rw-kernel AS TW BA

(e) Closeness C. (f) PageRank (g) Execution Time (sec)

Fig. 2. Similarity metrics comparison for TW dataset

118 T. Bakogiannis et al.

 Comparing Similarity Measures: The results of the similarity measure

comparisons, in the case of the TW dataset, are displayed in Fig. 2, where
M dAP E is used to express the modeling error. We compare six similarity mea-
sures: The degree distribution + levels measure (for levels equal from 0 to 2), a
combination of level-0 degree distribution with vertex count (denoted by level-
0 + size), D-measure and the Random Walk Kernel based similarity measure
(denoted by rw-kernel). The results indicate the impact that the choice of similar-
ity measure has on modeling accuracy. A more suitable to the modeled operator
and detailed similarity measure is more sensitive to topology differences and can
lead to better operator modeling.
In all Figures, with the exception of PageRank, we observe that the degree dis-
tribution + levels similarity measure, for a number of levels, can model an operator
more accurately than the simple degree distribution-based, effectively reducing the
errors reported in Table 2. Indeed, the addition of more levels to the degree distri-
bution incorporates more information about the connectivity of each vertex. This
additional topological insights contribute positively to better estimate the simi-
larity of two graphs. Examining the modeling quality, we observe that it increases
but only up to a certain point, in relation to the topology of the graphs in the
dataset. For example, since TW comprises of ego graphs, all the degrees of level
> 2 are zero, since there exist no vertices with distance greater than 2; therefore,
employing more levels does not contribute any additional information about the
topology of the graphs when computing their similarity. Finally, we observe that, in
specific cases, such as PageRank (Fig. 2f), enhancing the degree distribution with
degrees of more levels introduces information that is interpreted as noise during
modeling. PageRank is better modeled with the simple degree distribution as a
similarity measure. As such, we argue that for a given dataset and graph operator,
experimentation is required to find the number of levels that give the best trade-off
between accuracy and execution time.
We next concentrate on the effect of the combination of degree distribution
with vertex count in the modeling accuracy. We note that the vertex count
contributes positively in the modeling of distance-related metrics while having
a neutral or negative impact on degree- and spectrum-related metrics. This is
attributed to the existence of, at least, a mild correlation, between vertex count
and bc, ebc and cc [22]. For our least accurately approximated task, edge between-
ness centrality, employing the combination of measures results in a more than
6× decrease in error.
For D-measure, our experiments show that, for distance-related metrics it
performs at least as good as the degree distribution + levels similarity mea-
sures for a given level, with the notable exception of the PageRank case. On
the other hand, the degree distribution can be sufficiently accurate for degree-
or spectrum-related metrics. As D-measure is based on distance distributions
between vertices, having good accuracy for distance-related measures is some-
thing to be expected. A good example of the effectiveness of D-measure is
shown in the case of closeness centrality that involves all-pairs node distance
information directly incorporated in D-measure as we have seen in Sect. 3. In
 Predicting Graph Operator Output over Multiple Graphs 119

Fig. 2e we observe that by adding levels we get better results, vertex count con-
tributes into even better modeling but D-measure gives better approximations.
Yet, our methods’ errors are already very small (less than 3%) in this case. Con-
sidering the rw-kernel similarity measure, we observe that it performs poorly for
most of the operators. Although its modeling accuracy is comparable to degree
distribution + levels for some operators, we find that for a certain level or in
combination with vertex count a degree distribution-based measure has better
accuracy. Notably, rw-kernel has low accuracy for degree and distance related
operators while performing comparably in the case of spectrum operators.

0.04 0.12

0.1

0.08
MdAPE

MdAPE
0.02 0.06

0.04

0.02

0 0
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.05 0.1 0.15 0.2 0.25 0.3 0.35
Sampling Ratio Sampling Ratio
level-0 level-2 rw-kernel level-0 level-2
level-1 level-0+size level-1 level-0+size

(a) AS dataset (b) BA dataset

Fig. 3. Similarity Metric Comparison for Betweenness C.

Identifying betweenness centrality as one of the hardest operators to model

accurately, we present M dAP E approximation errors for AS and BA in Figs. 3a,
3b. These Figures do not include D-measure, since it was not possible to compute
it because of its running time. We note that the approximation error is below
12% and that the degree distribution + levels measures further improve on it for
both datasets. Compared to TW (Fig. 2), we observe that the level-2 similarity
measure provides better results for AS and BA but not TW, attributed to the
ego graph structure with level-2 degrees being zero. Finally, it is expected that
level-0 + size for BA to be no different than plain level-0, since all the graphs
in BA have the same vertex count by construction.
The aforementioned similarity measures have striking differences in their exe-
cution time. A comparison in computation time for different levels of the degree
distribution + levels similarity measure is presented in Fig. 2g. In the case of D-
measure, the actual execution time is presented for the TW dataset, since it was
prohibitively slow to compute it for the other two datasets. For the remaining two
datasets, we have computed D-measure on a random number of pairs of graphs
and then projected the mean computation time to the number of comparisons
performed by our method for each dataset.
Our results show that the overhead from level-0 to level-1 is comparable
for all the datasets. However, that is not the case for level-2. The higher the
level, the more influential the degree of the vertices becomes in the execution
time. Specifically, while we find level-0 to be 3.2× faster than level-2 for TW, we
120 T. Bakogiannis et al.

observe that in the case of AS and BA it is 19× and 76× faster. The computation
of the D-measure and the rw-kernel, on the other hand, are orders of magnitude
slower. Given the difference in modeling quality between the presented similarity
functions, we observe a clear trade-off between quality of results and execution
time in the context of our method.

4 Related Work
Our work relates to the actively researched areas of graph similarity, graph
analytics and machine learning. The available techniques for quantifying graph
similarity can be classified into three main categories ([24,36]):
 Graph Isomorphism - Edit Distance: Two graphs are considered similar
if they are isomorphic. A generalization of the graph isomorphism problem is
expressed through the Edit Distance, i.e., the number of operations that have to
be performed in order to transform one graph to the other [31]. The drawback
of approaches in this category is that graph isomorphism is hard to compute.
 Iterative Methods: This category of graph similarity algorithms is based on
the idea that two vertices are similar if their neighborhoods are similar. Applying
this idea iteratively over the entire graph can produce a global similarity score.
Such algorithms compare graphs based on their topology, we choose to map
graphs to feature vectors and compare those vectors instead.
 Feature Vectors: These approaches are based on the idea that similar graphs
share common properties such as degree distribution, diameter, etc and therefore
represent graphs as feature vectors. To assess the degree of similarity between
graphs, statistical tools are used to compare their feature vectors instead. Such
methods are not computationally demanding. Drawing from this category of
measures, we base our graph similarity computations on comparing degree dis-
tributions.
 Graph Kernels: A different approach to graph similarity comes from the
area of machine learning where kernel functions can be used to infer knowledge
about samples. Graph kernels are kernel functions constructed on graphs or
graph nodes for comparing graphs or nodes respectively. Extensive research on
this area (e.g., [15,17]) has resulted in many kernels based on walks, paths,
etc. While computationally more expensive they provide a good baseline for our
modeling accuracy evaluation.
 Graph Analytics and Machine Learning: Although graph analytics is a
very thoroughly researched area, there exist few cases where machine learning
techniques are used. On the subject of graph summarization, a new approach is
based on node representations that are learned automatically from the neighbor-
hood of a vertex [18]. Node representations are also applicable in computing node
or graph similarities as seen in [18]. However, we do not find works employing
machine learning techniques in the field of graph mining through graph topology
metric computations.
 Predicting Graph Operator Output over Multiple Graphs 121

5 Conclusion
In this work we present an operator-agnostic modeling methodology which lever-
ages similarity between graphs. This knowledge is used by a kNN classifier to
model a given operator allowing scientists to predict operator output for any
graph without having to actually execute the operator. We propose an intuitive,
yet powerful class of similarity measures that efficiently capture graph relations.
Our thorough evaluation indicates that modeling a variety of graph operators is
not only possible, but it can also provide results of high quality at considerable
speedups. Finally, our approach appears to present similar results to state-of-
the-art similarity measures, such as D-measure, in terms of quality, but requires
orders of magnitude less execution time.

References
1. Aherne, F.J., et al.: The bhattacharyya metric as an absolute similarity measure
for frequency coded data. Kybernetika 34(4), 363–368 (1998)
2. Arthur, D., Vassilvitskii, S.: k-means++: the advantages of careful seeding. In:
Proceedings of the Eighteenth Annual ACM-SIAM Symposium on Discrete Algo-
rithms, SODA 2007, New Orleans, Louisiana, USA, pp. 1027–10357–9 January
2007
3. Barabási, A.L., Albert, R.: Emergence of scaling in random networks. Science
286(5439), 509–512 (1999)
4. Bentley, J.L.: Multidimensional binary search trees used for associative searching.
Commun. ACM 18(9), 509–517 (1975)
5. Bhattacharyya, A.: On a measure of divergence between two statistical populations
defined by their probability distributions. Bull. Calcutta Math. Soc. 35(1), 99–109
(1943)
6. Bonacich, P.: Power and centrality: a family of measures. Am. J. Sociol. 92(5),
1170–1182 (1987)
7. Bounova, G., de Weck, O.: Overview of metrics and their correlation patterns for
multiple-metric topology analysis on heterogeneous graph ensembles. Phys. Rev.
E 85, 016117 (2012)
8. Brandes, U., Erlebach, T.: Network Analysis: Methodological Foundations.
Springer, New York (2005). https://doi.org/10.1007/b106453
9. Brin, S., Page, L.: The anatomy of a large-scale hypertextual web search engine.
Comput. Netw. 30(1–7), 107–117 (1998)
10. Csardi, G., Nepusz, T.: The Igraph software package for complex network research.
Inter J. Complex Syst. 1695, 1–9 (2006)
11. Eppstein, D., Wang, J.: Fast approximation of centrality. J. Graph Algorithms
Appl. 8, 39–45 (2004)
12. da, F., Costa, L., et al.: Characterization of complex networks: a survey of mea-
surements. Adv. Phys. 56(1), 167–242 (2007)
13. Freeman, L.C.: A set of measures of centrality based on betweenness. Sociometry
40(1), 35–41 (1977)
14. Gandomi, et al.: Beyond the hype. Int. J. Inf. Manage. 35(2), 137–144 (2015)
15. Gärtner, T.: A survey of kernels for structured data. SIGKDD 5(1), 49–58 (2003)
122 T. Bakogiannis et al.

16. Gärtner, T., Flach, P., Wrobel, S.: On graph kernels: hardness results and efficient
alternatives. In: Schölkopf, B., Warmuth, M.K. (eds.) COLT-Kernel 2003. LNCS
(LNAI), vol. 2777, pp. 129–143. Springer, Heidelberg (2003). https://doi.org/10.
1007/978-3-540-45167-9 11
17. Ghosh, S., Das, N., Gonçalves, T., Quaresma, P., Kundu, M.: The journey of graph
kernels through two decades. Comput. Sci. Rev. 27, 88–111 (2018)
18. Grover, A., Leskovec, J.: node2vec: scalable feature learning for networks. In:
SIGKDD, pp. 855–864. ACM (2016)
19. Hastie, T., Tibshirani, R., Friedman, J.: The Elements of Statistical Learning: Data
Mining, Inference, and Prediction, 2nd edn. Springer, New York (2009). https://
doi.org/10.1007/978-0-387-21606-5
20. Hernández, J.M., Mieghem, P.V.: Classification of graph metrics, pp. 1–20 (2011)
21. Jamakovic, A., et al.: Robustness of networks against viruses: the role of the spec-
tral radius. In: Symposium on Communications and Vehicular Technology, pp.
35–38 (2006)
22. Jamakovic, A., Uhlig, S.: On the relationships between topological measures in
real-world networks. NHM 3(2), 345–359 (2008)
23. Kaufmann, L., Rousseeuw, P.: Clustering by means of medoids, pp. 405–416 (1987)
24. Koutra, D., Parikh, A., Ramdas, A., Xiang, J.: Algorithms for graph similarity and
subgraph matching. Technical Report Carnegie-Mellon-University (2011). https://
people.eecs.berkeley.edu/∼aramdas/reports/DBreport.pdf
25. Leskovec, J., Adamic, L.A., Huberman, B.A.: The dynamics of viral marketing.
TWEB 1(1), 5 (2007)
26. Leskovec, J., Krevl, A.: SNAP Datasets: Stanford large network dataset collection,
June 2014. http://snap.stanford.edu/data
27. Leskovec, J., Sosič, R.: Snap: a general-purpose network analysis and graph-mining
library. ACM Trans. Intell. Syst. Technol. 8(1), 1 (2016)
28. McAuley, J.J., Leskovec, J.: Learning to discover social circles in ego networks. In:
Bartlett, P.L., Pereira, F.C.N., Burges, C.J.C., Bottou, L., Weinberger, K.Q. (eds.)
Advances in Neural Information Processing Systems 25: 26th Annual Conference
on Neural Information Processing Systems 2012. Proceedings of a meeting held
December 3–6, 2012, Lake Tahoe, Nevada, United States, pp. 548–556 (2012)
29. Newman, M.E.J., Girvan, M.: Finding and evaluating community structure in net-
works. Phys. Rev. E 69(2), 026113 (2004)
30. Riondato, M., Kornaropoulos, E.M.: Fast approximation of betweenness centrality
through sampling. Data Min. Knowl. Discov. 30(2), 438–475 (2016)
31. Sanfeliu, A., Fu, K.: A distance measure between attributed relational graphs for
pattern recognition. IEEE Trans. Syst. Man Cybern. 13(3), 353–362 (1983)
32. Schieber, T.A., et al.: Quantification of network structural dissimilarities. Nat.
Commun. 8, 13928 (2017)
33. Sugiyama, M., Borgwardt, K.M.: Halting in random walk kernels. In: Annual Con-
ference on Neural Information Processing Systems, pp. 1639–1647 (2015)
34. Bakogiannis, T., Giannakopoulos, I., Tsoumakos, D., Koziris, N.: Graph operator
modeling over large graph datasets. CoRR abs/1802.05536 (2018). http://arxiv.
org/abs/1802.05536
35. Vishwanathan, S.V.N., Schraudolph, N.N., Kondor, R., Borgwardt, K.M.: Graph
kernels. J. Mach. Learn. Res. 11, 1201–1242 (2010)
36. Zager, L.A., Verghese, G.C.: Graph similarity scoring and matching. Appl. Math.
Lett. 21(1), 86–94 (2008)