Explanation Retrieval in Semantic Networks
Understanding Spreading Activation based Recommendations
Vanessa N. Michalke and Kerstin Hartig
Daimler Center for Automotive IT Innovations (DCAITI), TU Berlin, Ernst-Reuter-Platz 7, Berlin, Germany
Keywords:
Explanation Retrieval, Spreading Activation, Pattern Recognition, Information Retrieval.
Abstract:
Spreading Activation is a well-known semantic search technique to determine the relevance of nodes in a
semantic network. When used for decision support, meaningful explanations of semantic search results are
crucial for the user’s acceptance and trust. Usually, explanations are generated based on the original network.
Indeed, the data accumulated during the spreading activation process contains semantically extremely valuable
information. Therefore, our approach exploits the so-called spread graph, a specific data structure that com-
prises the spreading progress data. In this paper, we present a three-step explanation retrieval method based on
spread graphs. We show how to retrieve the most relevant parts of a network by minimization and extraction
techniques and formulate meaningful explanations. The evaluation of the approach is then performed with a
prototypical decision support system for automotive safety analyses.
1 INTRODUCTION
Recommender, advisory, and expert systems utilize
available domain information or knowledge, often
in order to help guide decision makers in decision-
making processes (Kaklauskas, 2015). Semantic or
associative networks may be used for representing
such complex knowledge. Spreading activation algo-
rithms support searching those nets semantically, pro-
viding information about relevance of specific nodes
with respect to the specified search goals. Spreading
activation algorithms base upon the wavelike distribu-
tion of activation values from nodes to neighbor nodes
throughout the searched network. Consequently, a
node’s level of activation reveals its level of relevance
in the current semantic search.
Semantic search results, such as the ones retrieved
by spreading activation, can be used for providing rec-
ommendations and advices in decision support sys-
tems in various domains, e.g., medical diagnostics
(Alvarez et al., 2011), or automotive safety analyses
(Hartig and Karbe, 2016). In many applications, plau-
sibility and understandability of the search results be-
come important for users, i.e., the decision-makers,
because they need to understand why a certain recom-
mendation or advice was provided. Therefore, expla-
nations of the results are crucial for the user’s accep-
tance and trust in decision support systems. Current
approaches usually identify relevant paths and envi-
ronments by exploiting the original network. This
neglects the valuable ancillary information accumu-
lated during the search process. In this paper, we
present a graph-based method for automatically gen-
erating explanations for semantic search results re-
trieved by applying spreading activation techniques.
Our approach attempts to identify the nodes and edges
that contributed most to the search result we try to ex-
plain. Therefore, we retrieve the most relevant extract
not from the underlying semantic network but from
a specific graph structure representing the spreading
activation progress during search, i.e., the so-called
spread graph.
The remainder of this paper is structured as fol-
lows. In Section 2, we examine related work in the
areas of spreading activation, recommender systems
and explaining information retrieval results. In Sec-
tion 3, we present our three-step method for the au-
tomated generation of explanations. In Section 4, the
fulfillment of explanation goals and the goodness of
the generated explanations will be assessed. Finally,
in Section 5, we summarize our results and present
multiple possibilities to continue research.
2 RELATED WORK
In this section, we introduce the basics of searching
semantic networks by spreading activation and its ap-
Michalke, V. and Hartig, K.
Explanation Retrieval in Semantic Networks - Understanding Spreading Activation based Recommendations.
DOI: 10.5220/0006050502910298
In Proceedings of the 8th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2016) - Volume 1: KDIR, pages 291-298
ISBN: 978-989-758-203-5
Copyright
c
2016 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
291
plication in recommender systems. We then outline
recent research in the area of explanation generation
as a meaningful extension for such systems.
2.1 Spreading Activation
One possibility for identifying and ranking relevant
regions in semantic networks is spreading activa-
tion. In a comprehensive survey, Crestani provides
an overview of its various applications for associa-
tive information retrieval (Crestani, 1997). Techni-
cally, starting from at least one start node, activation
is passed to connected nodes. Thereby, the network is
flooded wave-like, in so-called pulses, until a termi-
nation condition is satisfied. The spreading process
is controlled by specific activation functions, allow-
ing the calculation of different activation values, such
as input and output activation as well as the activa-
tion level of each node. In each pulse, the activation
value a node received is referred to as its input acti-
vation whereas the activation value distributed by a
node is referred to as its output activation. Input ac-
tivation combines activation values potentially trans-
ported by several edges. Therefore, the term trans-
ported activation of an edge corresponds to its actual
activation value passed in a given pulse. The activa-
tion level of a node represents its current relevance
level in the search. Thus, after applying spreading
activation, relevant regions of a network can be iden-
tified, enabling application-specific interpretation.
2.2 Recommendations
Today, web-based recommender systems are well-
established, e.g., providing useful suggestions for
products or videos (Jain et al., 2015). One system
that recommends items by using spreading activation
is WebSCSA (Web Search by Constrained Spread-
ing Activation), as described in (Crestani and Lee,
2000). WebSCSA identifies relevant web pages based
on marked pages during a user’s web search. Besides
web-based applications, recommendations in advi-
sory systems support the management of complex
domain-specific knowledge in enterprises. Such advi-
sory systems may support decision-makers in diverse
domains, e.g., medical diagnostics (Alvarez et al.,
2011), or automotive safety analyses (Hartig and
Karbe, 2016). The importance of high-quality deci-
sions makes explanations especially valuable. There-
fore, such systems are potential candidates for the ex-
planation retrieval method presented in this paper.
2.3 Explanations
The user needs to understand the reasons for a spe-
cific recommendation to assess its quality (Sinha and
Swearingen, 2002). Explanations can support this un-
derstanding. Reconstructive explanation generation,
as described in (Wick and Thompson, 1992), is con-
sidered an active problem-solving process to explain
any results received by a problem solving component.
Based on the problem solving computations, i.e., the
so-called line-of-reasoning, it generates a so-called
line-of-explanation. For this purpose, the prototype
REX (Reconstructive Explanation System) uses the
shortest path identified by the A* algorithm (Wick
and Thompson, 1992). The method, presented in this
paper follows the general concepts of reconstructive
explanation generation, but focuses on spreading ac-
tivation as one particular problem solving technique.
In (Forcher et al., 2010), explanations are also cre-
ated by using the shortest path. However, other cri-
teria than the path length must be considered when
searching for the best explanation. (Aleman-Meza
et al., 2005) rank the relationships in semantic net-
works by using configurable semantic and statistical
criteria like subsumption or popularity. Similar crite-
ria are used for the assessment of relevant paths, iden-
tified by a modified bidirectional breadth-first search
(Viswanathan and Krishnamurthi, 2012) and a seman-
tic ant colony optimization algorithm (Viswanathan
and Krishnamurthi, 2015). When explaining spread-
ing activation results, an assessment by the presented
criteria is generally applicable. However, we assume
that not limiting explanation retrieval to paths can be
of greater value for meaningful explanations.
In contrast to path-based approaches, subgraph-
based approaches consider potentially relevant ele-
ments across the entire network. Many subgraph-
based approaches extract connection subgraphs from
the original graph and utilize it for their explanation
deduction (Faloutsos et al., 2004; Forcher et al., 2011;
Forcher et al., 2012). Connection subgraphs aim at
describing the relationships between nodes after par-
ticular relevance analyses, e.g., including the appli-
cation of goodness criteria, keyword assignment, or
neighborhood expansion. However, these approaches
utilize the original graph for relevance assessment.
Since semantic search by spreading activation is a
means of relevance assessment of mutual nodes, we
suppose the answer to the question why a node is
highly relevant to another to ground on the spread-
ing activation process itself. Therefore, we propose to
utilize the valuable spreading activation data, i.e., our
spread graph, a specific graph-based structure repre-
senting logged spreading activation process data.
KDIR 2016 - 8th International Conference on Knowledge Discovery and Information Retrieval
292
3 METHOD FOR EXPLAINING
RECOMMENDATIONS
In this paper, we present a method for explaining
spreading activation based recommendations, which
takes advantage of available additional information
from the spreading activation process itself. The
method aims at explaining the reason why a recom-
mendation, i.e., a result node in the searched semantic
network, is considered to be of relevance for the given
start nodes of the spreading activation process. There-
fore, we utilize the data produced during spreading
activation and represent it in a dedicated spread graph
structure. A spread graph is a directed graph illus-
trating the spreading activation process on the original
graph. Nodes in the spread graph represent the state of
their corresponding original graph’s nodes in a certain
spreading pulse. Each edge represents one activation
value distribution step via the corresponding original
graph’s edge. The source(s) of a spread graph are the
start node(s) of the corresponding spreading process.
Figure 1 depicts the transformation from the orig-
inal graph to the spread graph as well as two tables il-
lustrating the performed spreading activation steps. In
the left table, each row describes one activation distri-
bution from one node to another in a pulse p including
output o
f rom,p
and transported activation t
edge,p
. The
right table shows the input activation i
v,p
and activa-
tion values a
v,p
for each node after each pulse. The
spread graph combines both the elements from the
original graph and the unrolled spreading activation
steps. Therefore, the spread graph may contain more
than one corresponding element to one original ele-
ment, e.g., v
0
1
and v
00
1
correspond to v
1
.
Original graph and spreading activation steps
v
1
v
5
e
1
v
2
v
4
v
3
e
3
e
4
e
2
v
1
v
2
v
3
v
4
‘‘
v
1
‘‘
e
2
e
1
e
1
‘‘
e
3
Spread graph
p
from
to
via
o
from,p
t
edge,p
1
v
1
v
2
e
1
o
v
1
,1
t
e
1
,1
1
v
1
v
3
e
2
o
v
1
,1
t
e
2
,1
2
v
2
v
1
e
1
o
v
2
,2
t
e
1
,2
2
v
3
v
4
e
3
o
v
3
,2
t
e
3
,2
...
p
v
v,p
a
v,p
1
v
2
v
2
a
v
2
,1
1
v
3
v
3
a
v
3
,1
2
v
1
v
1
a
v
1
,2
2
v
4
v
4
a
v
4
,2
i
v
1
,2
a
v
1
,2
p=2
t
e
1
,2
i
v
2
,1
a
v
2
,1
o
v
2
,2
p=1
t
e
1
,1
p=1
t
e
2
,1
a
v
1
,0
o
v
1
,1
T
R
A
N
S
F
O
R
M
A
T
I
O
N
Figure 1: Transformation to spread graph.
Knowledge-based recommender systems often
utilize ontologies for knowledge representation, for-
mulated in different RDF-based target languages,
from plain RDF to the more expressive Web Ontol-
ogy Language. However, they all base on an underly-
ing RDF Graph, which can be visualized as directed
graph, where the contained statements are represented
by nodes and directed edges (RDF, 2014). The ap-
proach presented in this paper is applicable to any
target structure based on RDF. Usually, a recommen-
dation retrieved by spreading activation is character-
ized by the highest activation value. However, since
the presented approach is capable of explaining every
spreading activation result, the term result node spec-
ifies the node an explanation is required for.
Step 1
Minimization
of the
Problem Area
Step 2
Identification
of
Relevant Groups
Step 3
Pattern Recognition
and Explanation
Creation
Explanation
Recommen-
dation
Generation
Figure 2: Method overview.
Figure 2 illustrates our three-step method for the
explanation of a given recommendation. Initially, we
apply a minimization technique to remove all nodes
from the spread graph that are strictly irrelevant for
the explanation. On base of the resulting minimized
spread graph the most relevant subgraphs, so-called
groups, are identified in a second step by using com-
plex relevance detection techniques. In a third step,
group-based explanations are created using pattern
recognition. In the following sections, we provide a
detailed description of each step together with a con-
tinuous example.
5
4
16
0
1
7
3
12
0
1
6
3
10
7
11
Pulse 1 Pulse 2 Pulse 3
2
8
13
14
16
5
4
12
0
1
6
3
10
7
11
8
13
14
16
9
17
18
start node
result node
Figure 3: Example spread graph.
Example: In Figure 3, an example spread graph is
depicted representing the states of each node at each
pulse in a three-pulse spreading activation process.
For example, node 3 receives activation from node 2
in the first pulse. The objective of this work is to ex-
plain the detected high relevance of result node 11 for
start node 2. The corresponding original graph and
the remaining spreading data are not depicted since
they are not required for understanding the example.
Explanation Retrieval in Semantic Networks - Understanding Spreading Activation based Recommendations
293
3.1 Minimization of the Problem Area
Since a spread graph represents the entire spreading
activation process, it may contain unnecessary infor-
mation for the explanation of concrete result nodes.
Nodes and edges in the spread graph that do not con-
tribute directly or indirectly to the activation of a re-
sult node, we consider to be strictly irrelevant. This
applies to all nodes and edges that are not part of any
path between start and result nodes. We refer to them
as dead-end elements and propose their removal in or-
der to minimize the problem area. As a result, we re-
trieve a minimized spread graph containing only po-
tentially relevant nodes and edges. Figure 4 illustrates
various dead-end elements, depicted as dashed lines.
start
node
result
node
dead-end
elements
potentially
relevant
elements
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Figure 4: Minimization by backtracking.
Technically, the minimization follows the idea of
impact analysis. Starting at the result nodes, all paths
are tracked back in the edges’ opposite direction until
any start node is reached. All elements that are not
part of the detected tracks are dead-end elements and
can be ignored safely. The paths of the backtrack-
ing form the resulting minimized spread graph. Dur-
ing minimization, the activation values of the spread
graph are transferred to the corresponding elements in
the minimized spread graph.
Example: Figure 5 shows the result of minimization
by dead-end removal for the spread graph in Figure 3.
The minimized spread graph only contains potentially
relevant information for the explanation of the result
node 11. For example, node 8 (located between pulse
two and three) is not part of any path that starts at start
node 2 and ends at any result node 11. That means,
node 8 did not contribute anything to the relevance of
node 11 and can be removed.
16
11
Pulse 3
12
6
16
0
1
7
3
Pulse 2 Pulse 1
2
start node
result node
Figure 5: Example minimized spread graph.
3.2 Identification of Relevant Groups
Although a minimized spread graph is usually smaller
than its corresponding spread graph, it still tends to be
too large in order to create a meaningful and concise
textual explanation. Therefore, the goal of this step is
the identification of the most relevant subgraphs, so-
called groups, of a minimized spread graph. Their
detection strongly relies on the analysis of previ-
ously collected activation values assigned to the (min-
imized) spread graph. This data allows conclusions
about the relevance of nodes and edges at different
pulses. To identify such groups, we present two sub-
steps: selecting the most relevant nodes and determin-
ing the most relevant paths and groups around them.
3.2.1 The Most Relevant Nodes
Since each node in the minimized spread graph car-
ries a certain relevance, considering all of them in an
explanation poses a very complex task. In order to re-
duce complexity, we introduce the additional parame-
ter n > 0 to define the expected number of most rele-
vant nodes. The precision n can be configured by a
user or, alternatively, can be automatically determined
based on specific properties of the given minimized
spread graph, e.g., the number of represented pulses.
Potential candidates for most relevant nodes are
all nodes in a minimized spread graph that are neither
start nodes, result nodes, or semantically meaningless
nodes such as class or property nodes in RDF-based
semantic networks. Additionally, an application-
specific filtering of the candidate nodes set is possible.
The relevance of a candidate node is indicated by its
relevance value r. The calculation of r is based on its
input activation because it composes all transported
activation values to the candidate in the given pulse
and, therefore, denotes its relevance. Furthermore,
two properties of the surrounding local network struc-
ture, i.e., connectivity and existence of class nodes,
are used for the calculation. In the following, both are
introduced in detail.
First, class nodes only represent the type of an in-
stance node, for example Ella is of type canary. A
name of an instance can be used together with its
type to raise the comprehensibility of an explanation.
However, types do not add further meaningful infor-
mation to explanations. For that reason, the trans-
ported activation from these nodes to the candidates
is subtracted from the input activation of the candi-
dates, which leads to a reduced relevance value. Ad-
ditionally, the relevance of the candidates that activate
class nodes is slightly reduced to take into account
their special relationship to class nodes.
Second, the connectivity of a candidate influences
its relevance. A strong distribution of activation is an
indicator for a very broad semantic meaning (Crestani
and Lee, 2000) and, therefore, indicates less rele-
vance for an explanation. On the other side, highly
connected nodes, which receive activation over many
edges, tend to get high input activation values. Never-
KDIR 2016 - 8th International Conference on Knowledge Discovery and Information Retrieval
294
theless, candidates with lower input activation values
might be more relevant for an explanation, e.g., when
the activating edges carried more semantic meaning.
In Figure 6(a), node x receives activation by only two
nodes, whereas in Figure 6(b), node y receives acti-
vation by many nodes. The overall input activation
i of node y exceeds the input activation of node x
by 1. However, node x should be considered to be
more relevant for an explanation than node y. In order
to compare highly connected nodes with other nodes,
the ratio of in- and outgoing edges is considered dur-
ing relevance value computation.
x
v
2
v
1
y
v
1001
v
1
t=1
t=1
t=500
t=500
i=1000
i=1001
(b) (a)
Figure 6: Highly connected nodes.
The relevance of each node is calculated by de-
fined relevance functions based on its input activation
with respect to the beforementioned properties. Then,
the n nodes with the highest relevance values from the
candidate set represent the most relevant nodes.
Example: The input activation and the calculated
relevance values for all candidates in the minimized
spread graph are shown in Figure 7. Start node 2 and
result node 11 are by definition no candidates as well
as the class nodes 16. Nodes 12 and 3 are the resulting
relevant nodes due to their highest relevance values.
Note, that node 6 has a higher input activation than
node 3. However, regarding the relevance value, node
3 outperforms node 6 because of the connectivity.
relevance value
16
11
Pulse 3
12
6
16
0
1
7
3
Pulse 2 Pulse 1
3,8475 | 0,7695
11,34 | 2,268
1,35 | 0,45
1,35 | 0,45
1,8 | 0,9
1,35 | 0,45
class node
2
input activation
start node
result node
Figure 7: Weighted example minimized spread graph.
3.2.2 The Most Relevant Paths and Groups
The detected most relevant nodes do not form an ex-
planation yet due to missing information about their
relationship to the start and result node(s). This re-
lationship can be clarified through the paths leading
from the result node via a relevant node to a start node
in the minimized spread graph. The problem of find-
ing such paths can be transformed into the two prob-
lems of finding a set of paths between the result and
relevant node and between the relevant and the start
node(s), as shown in Figure 8. We define an additional
parameter m > 0 for each subproblem, which restricts
the number of paths to identify in order to reduce the
explanation length. This value is either user-specified
or based on specific properties.
result
node
relevant
node
relevant
node
start
node
result
node
relevant
node
start
node
result group
start group
Figure 8: Dividing the path-finding problem.
For the explanation, we search for the most ex-
pressive and, therefore, the most relevant paths. Since
the transported activation reflects the relevance of
an edge, the relevance of a path can be determined
by the sum of transported activation values along its
edges. Thus, the most relevant paths are the longest
paths regarding their overall transported activation.
Generally, finding a longest path is NP-hard (Uehara
and Uno, 2004). However, directed acyclic graphs,
such as (minimized) spread graphs, allow an effi-
cient solution of the problem. In (Sedgewick and
Wayne, 2011), the authors propose the determination
of longest paths by using a shortest path algorithm
with negated edge weights. We apply this approach
to a Bellman-Ford algorithm. In contrast to origi-
nal graphs, spread graphs support the efficient longest
path search. This is an advantage of the approach pre-
sented in this paper.
Each identified set of most relevant paths for one
relevant node can be represented as a subgraph of
the minimized spread graph. We call such subgraphs
groups, whereas the group from a result node to a rele-
vant node is called result group and the group from a
relevant node to start node(s) is called start group (see
Figure 8). Distinguishing between these two groups is
a benefit, because it allows customized explanations
for their diverging meaning. While the start group
explains the origin of a relevant node, i.e., the recom-
mendation request, the result group explains its effect,
i.e., recommendation.
Example: Based on the minimized spread graph, the
start and result groups for the relevant nodes 12 and
3 are illustrated in Figure 9. For this example, the
maximum number of paths per group (parameter m)
is restricted to 2. Therefore, the start group of node 12
does not contain all potentially relevant paths.
result group
start group
result group
start group
3
12
0
3
11 12
11
12 3
2
2
start node
result node
relevant node
Figure 9: Example groups for the relevant nodes.
Explanation Retrieval in Semantic Networks - Understanding Spreading Activation based Recommendations
295
3.3 Pattern Recognition and
Explanation Creation
The benefit of group-based explanations is that we do
not depend on explaining paths as one-dimensional
statements but can take advantage of cross-path in-
formation within the graph. Pattern analysis allows to
linguistically subsume elements in an explanation that
structurally belong together in the underlying group.
This prevents confusing explanations and supports
conciseness, compactness, and comprehensibility.
v
i
v
v
1
v
i
v
v
1
v
2
v
v
1
v
i
v
v
1
v
e
1
e
1
e
1
e
1
e
1
e
i
e
i
e
i
e
i
e
2
v
v
1
e
1
No pattern
Multiple outgoing edges
Multiple incoming edges
Chain of three
Diamond
low
priority
high
priority
level level level
Figure 10: Patterns.
We identified the patterns in Figure 10, which are
prioritized regarding their expressiveness: from low
priority for no specific pattern to highly prioritized di-
amonds. If no special pattern can be identified, two
nodes and their connecting edge can be interpreted as
a simple statement of subject, predicate, and object, as
common for semantic networks. In case one node has
at least two outgoing edges to other nodes, the pattern
of multiple outgoing edges applies. Similarly, there is
a pattern for multiple incoming edges. The chain of
three pattern applies whenever there is a sequence of
three nodes connected by unidirectional edges. The
diamond pattern describes two or more equally long
downward paths that split in the first node and rejoin
in the last node. However, the level of depth must be
exactly 3 and the nodes of a level and the edges be-
tween the same levels have to be of the same type.
A less restrictive concept for diamond structures was
introduced by (Furnas and Zacks, 1994) in the area
of multi-tree analysis. Since all other patterns are in-
cluded in a diamond, it contains most structural infor-
mation. This makes diamond structures especially ex-
pressive and valuable for explanations. After the pro-
posed pattern recognition, the actual sentence build-
ing process is straightforward. In this work, English
language is used. However, language-specific appli-
cation is possible. Sentence building follows two ba-
sic principles. First, as many nodes and edges as pos-
sible, that are located on the same depth level in the
pattern, should be subsumed, e.g., utilizing connec-
tion words such as and or using types. Second, the
different levels of a pattern need to be concatenated
to form a sentence, e.g., utilizing connection words
such as which. The generated textual explanation sup-
ports the user to comprehend the reasons for a given
spreading activation based recommendation.
Example: Figure 11 contains a sentence building ex-
ample for the described diamond pattern, which al-
lows to build only one short and more expressive sen-
tence instead of six sentences.
likes
to
cub B
canary
Ella
cub A
drink
water
has
likes to
The canary Ella has cub A. The canary Ella has
cub B. The canary Ella has cub C. Cub A likes to
drink water. Cub B likes to drink water. Cub C
likes to drink water.
cub C
has
has
likes to
The canary Ella has cubs, who like to drink water.
Figure 11: Sentence building example based on diamonds.
4 EVALUATION
Evaluating explanations is very challenging since the
most interesting evaluation criterion refers to useful-
ness. Usefulness can only be assessed subjectively,
mostly depending on the experience and preferences
of the observer (Klahold, 2009). In this paper, two
different evaluation approaches are utilized to show
the explanations’ usefulness. The first approach ar-
gues on the fulfillment of selected explanation goals
as proposed by (Tintarev and Masthoff, 2011). The
second approach evaluates the goodness of explana-
tions by estimating precision, recall, and F-measure.
These measures are common for assessing the good-
ness of recommendations and are applicable to expla-
nations as well (Klahold, 2009).
4.1 Fulfillment of the Explanation Goals
In (Tintarev and Masthoff, 2011), the authors propose
an explanation goal dependent evaluation of explana-
tions with goals like transparency, effectiveness, and
efficiency. Furthermore, they provide guidelines for
the application-specific selection of goals and their
evaluation. In this work, we focus on examining the
goals transparency as well as the support of efficient
and effective decisions for users.
Transparency: The proposed explanation generation
process follows a strict and reproducible process, uti-
lizing different intermediate graph structures. Each
intermediate graph is based on its predecessor and
their mutual relationships are defined by transforma-
tion rules. Therefore, an explanation is traceable
backwards through the intermediate graph structures
to the spread graph depicting the spreading activation
process. The result node as well as its corresponding
nodes can be identified in each single process step.
KDIR 2016 - 8th International Conference on Knowledge Discovery and Information Retrieval
296
This supports the user’s understanding of how the sys-
tem works. We therefore consider the proposed expla-
nation generation process to be transparent.
Effectiveness and Efficiency: Usually, original
graphs contain much information and spread graphs
tend to be even larger. For example, an original graph
examined in this paper with more than 26 000 nodes
and 138 000 edges leads to a spread graph of 108 000
nodes and 136 000 edges after 6 pulses. Manual trace-
ability of this large amount of data is infeasible. In
contrast, the minimized spread graph only contains 90
nodes and 190 edges, the groups even consist of only
11 nodes and 10 edges. That means, the user retrieves
an immensely reduced data set, which supports an ef-
fective and efficient decision making process.
These arguments show that the generated explana-
tions are adequate to their purpose. The explanations
are transparent and support effective and efficient de-
cisions, which contributes to their usefulness.
4.2 Goodness of the Explanations
To evaluate the explanations’ goodness, we apply the
proposed explanation generation method to results
from the HARvESTer recommender system, which
supports safety experts when performing automotive
safety analyses (Hartig and Karbe, 2016). There,
semantic networks represent real data of completed
automotive safety analyses by an automobile manu-
facturer. For each node and edge of such networks,
the information whether or not it is expected to be
part of an explanation for a given recommendation
is assigned
1
. Then, an explanation is generated for
the recommendation using the proposed explanation
method. Both results can then be compared regarding
their precision, recall, and calculated F-measure. For
the evaluation, we examine three different requests,
e.g., inquiring malfunctions based on user-given func-
tions, with 4 to 6 pulses. They are performed on four
semantic networks of different sizes listed in Table 1.
Table 1: Sizes of examined networks.
Networks A B C D
# nodes 1800 775 1218 26432
# edges 8779 3426 5319 138773
The precision for each request and network is
shown in Figure 12 (left), grouped by nodes (light
gray) and edges (dark gray). The precision results
range from 33% to 100%, where 100% means that
1
Remark: The assignment of the expected values are de-
termined by one advanced user. Certainly, the evaluations’
reliability can benefit from an assignment by a safety expert
group. For this work, an expert group is not available. This
is planned for extensive future case studies.
the generated explanation contains no false positive
element, i.e., unexpected node or edge. The lower
precision rates, e.g., for network B, are caused by two
main reasons. First, multiple nodes and edges are of-
ten of similar importance. Second, the size of the ex-
planation is restricted by the presented parameters n
and m. Thus, it is possible that only some of several
nodes and edges of similar importance are chosen for
an explanation while other expected nodes and edges
are missing. This explains that receiving 100% is not
always possible. The recall for each request and net-
work is shown in Figure 12 (right). Results range
from 60% to 100%, where 100% means that all ex-
pected nodes or edges are part of a generated expla-
nation, e.g., for request 1 on network C. Lower recall
results are mainly caused by similar important nodes
and edges, resulting in similar good explanations.
Precision
25%
50%
75%
100%
Request
Network
1
A
2
3
1
B
2
3
1
C
2
3
1
D
2
3
Nodes
Edges
Recall
25%
50%
75%
100%
1
A
2
3
1
B
2
3
1
C
2
3
1
D
2
3
Nodes
Edges
1
Figure 12: Precision and recall for nodes and edges.
The quality of an explanation benefits more from
a high rate of expected elements in the explanation
than it loses by additional nodes and edges. There-
fore, recall can be considered to be more important
than precision, i.e., represented by a F
2
-measure. Its
results range between 62% and 100% with an aver-
age of 84% for nodes and 78% for edges (exclud-
ing outliers). The evaluation shows, especially under-
lined by the high average F-measure, that the content
of the generated explanation generally corresponds to
the expectations. Besides the explanation generation
method, the goodness of an explanation strongly de-
pends on the different user preferences, the expres-
siveness of the underlying knowledge base and the
configuration of the spreading activation algorithm.
5 CONCLUSION AND OUTLOOK
In this paper, we presented a method to retrieve ex-
planations for semantic search results obtained by
spreading activation techniques. The novelty of our
approach is the excessive exploitation of the data ac-
cumulated during the search process instead of uti-
lizing the original network. Therefore, our method
is performed on a specific graph representing the
spreading process, the so-called spread graph. We
presented our explanation retrieval method in three
steps. Firstly, we introduced a minimization tech-
Explanation Retrieval in Semantic Networks - Understanding Spreading Activation based Recommendations
297
nique for the spread graph in order to remove strictly
irrelevant parts. Secondly, we presented the retrieval
of the most relevant subgraphs, so-called groups,
from the minimized spread graph. Lastly, we de-
scribed pattern recognition techniques to facilitate
meaningful and concise explanation verbalization in
natural language. The evaluation of the approach
showed promising results. For the examined seman-
tic networks, we were able to highlight both the ful-
fillment of the explanation goals and the goodness of
the generated explanations. However, we see much
potential for future research. Since the identifica-
tion of relevant parts in the spread graph is based
on very complex relations, our approach can be re-
fined, e.g., by an additional consideration of more
complex neighborhood influences. A more extensive
pattern analysis can improve the conciseness of gener-
ated explanations. We furthermore plan on extended
case studies in real-world application environments to
see the benefit of the provided explanations, e.g., the
increasing trust in recommendations. Especially, an
explanation goodness comparison with existing ap-
proaches can emphasize the benefit of our approach.
REFERENCES
Aleman-Meza, B., Halaschek-Weiner, C., Arpinar, I. B.,
Ramakrishnan, C., and Sheth, A. (2005). Ranking
complex relationships on the semantic web. IEEE In-
ternet Computing, 9(3):37–44.
Alvarez, J. M., Polo, L., Jimenez, W., Abella, P., and Labra,
J. E. (2011). Application of the spreading activation
technique for recommending concepts of well-known
ontologies in medical systems. Proceedings of the 2nd
ACM Conference on Bioinformatics, Computational
Biology and Biomedicine (BCB ’11), page 626.
Crestani, F. (1997). Application of spreading activation
techniques in information retrieval. Artificial Intelli-
gence Review, 11(6):453–482.
Crestani, F. and Lee, P. L. (2000). Searching the web by
constrained spreading activation. Information Pro-
cessing and Management: an International Journal
- Artificial Intelligence and Information Retrieval,
36(4):585–605.
Faloutsos, C., McCurley, K. S., and Tomkins, A. (2004).
Fast discovery of connection subgraphs. In Proceed-
ings of the tenth ACM SIGKDD international confer-
ence on Knowledge discovery and data mining (KDD
’04), pages 118–127, New York. ACM.
Forcher, B., Agne, S., Dengel, A., Gillmann, M., and Roth-
Berghofer, T. (2012). Towards Understandable Ex-
planations for Document Analysis Systems. In 10th
International Workshop on Document Analysis Sys-
tems(DAS) IAPR, pages 6–10, Gold Cost.
Forcher, B., Roth-Berghofer, T., Sintek, M., and Dengel,
A. (2010). Constructing Understandable Explana-
tions for Semantic Search Results. In 17th Interna-
tional Conference, Knowledge Engineering and Man-
agement by the Masses (EKAW), pages 493–502, Lis-
bon. Springer.
Forcher, B., Roth-Berghofer, T., Sintek, M., and Dengel,
A. (2011). Semantic Logging: Towards Explanation-
Aware DAS. In International Conference on Doc-
ument Analysis and Recognition (ICDAR’11), pages
1140–1144, Beijing. IEEE.
Furnas, G. W. and Zacks, J. (1994). Multitrees: Enriching
and Reusing Hierarchical Structure. In Conference
on Human Factors in Computing Systems (CHI’94),
pages 330–336, Bosten, Massachusetts. ACM.
Hartig, K. and Karbe, T. (2016). Recommendation-
based Decision Support for Hazard Analysis and Risk
Assessment. In 8th International Conference on
Information, Process, and Knowledge Management
(eKNOW ’16), pages 108–111.
Jain, S., Grover, A., Thakur, P. S., and Choudhary, S. K.
(2015). Trends, Problems And Solutions of Recom-
mender System. In International Conference on Com-
puting, Communication and Automation (ICCCA’15),
pages 955–958, Noida. IEEE.
Kaklauskas, A. (2015). Intelligent Decision Support Sys-
tems. In Biometric and Intelligent Decision Making
Support, volume 81, pages 195–220. Springer.
Klahold, A. (2009). Empfehlungssysteme: Recommender
Systems - Grundlagen, Konzepte und L
¨
osungen [Rec-
ommender Systems - Fundamentals, concepts and so-
lutions]. Vieweg + Teubner, Wiesbaden.
RDF (2014). RDF 1.1 Concepts and Abstract Syntax. World
Wide Web Consortium - W3C. Accessed: 15.06.2016.
Sedgewick, R. and Wayne, K. (2011). Algorithms - Fourth
Edition. Addison-Wesley, Boston, Massachusetts.
Sinha, R. and Swearingen, K. (2002). The role of trans-
parency in recommender systems. In Extended Ab-
stracts on Human Factors in Computing Systems (CHI
EA, CHI), pages 830–831, New York. ACM.
Tintarev, N. and Masthoff, J. (2011). Designing and evaluat-
ing explanations for recommender systems. In Ricci,
F., Rokach, L., Shapira, B., and Kantor, P. B., ed-
itors, Recommender Systems Handbook, pages 479–
510. Springer, US.
Uehara, R. and Uno, Y. (2004). Efficient Algorithms for
the Longest Path Problem. In Algorithms and Compu-
tation - 15th International Symposium (ISAAC ’04),
pages 871–883, Hong Kong. Springer.
Viswanathan, V. and Krishnamurthi, I. (2012). Finding rele-
vant semantic association paths through user-specific
intermediate entities. Human-centric Computing and
Information Sciences, 2(9).
Viswanathan, V. and Krishnamurthi, I. (2015). Finding
relevant semantic association paths using semantic ant
colony optimization algorithm. Soft Computing - A
Fusion of Foundations, Methodologies and Applica-
tions, 19(1):251–260.
Wick, M. R. and Thompson, W. B. (1992). Reconstruc-
tive expert system explanation. Artificial Intelligence,
54(1-2):33–70.
KDIR 2016 - 8th International Conference on Knowledge Discovery and Information Retrieval
298