Music Event Detection Leveraging Feature Selection based on Ant Colony Optimization

Abstract

Announcements of events are regularly spread using the Internet, e.g., via online newspapers or social media. Often, these events involve playing music publicly that is protected by international copyright laws. Authorities entrusted with the protection of the artists' interests have to find unregistered music events in order to fully exercise their duty. As a requirement, they need to find texts in the Internet that are related to such events like announcements or reports. However, event detection is a challenging task in the field of Text Mining due to the enormous variety of information that needs to be considered and the large amount of data that needs to be processed. In this paper, a process chain for the detection of music events incorporating external knowledge is proposed. Furthermore, a feature selection algorithm based on ant colony optimization to find featurse with a high degree of explanatory power is presented. Finally, the performance of five different machine learning algorithms including two learning ensembles is compared using various feature sets and two different datasets. The best performances reach an F1-measure of 0.95 for music texts and 0.968 for music event texts, respectively.

Full text

Music Event Detection Leveraging Feature Selection based on Ant Colony Optimization
Jian Xi∗†, Michael Spranger†and Dirk Labudde†‡
†University of Applied Sciences Mittweida
Forensic Science Investigation Lab (FoSIL), Germany
Email: {xi, spranger}@hs-mittweida.de ‡Fraunhofer
Cyber Security
Darmstadt, Germany
Email: labudde@hs-mittweida.de
Abstract—Announcements of events are regularly spread using
the Internet, e.g., via online newspapers or social media. Often,
these events involve playing music publicly that is protected
by international copyright laws. Authorities entrusted with the
protection of the artists’ interests have to find unregistered music
events in order to fully exercise their duty. As a requirement,
they need to find texts in the Internet that are related to such
events like announcements or reports. However, event detection
is a challenging task in the field of Text Mining due to the
enormous variety of information that needs to be considered
and the large amount of data that needs to be processed. In
this paper, a process chain for the detection of music events
incorporating external knowledge is proposed. Furthermore, a
feature selection algorithm based on ant colony optimization
to find featurse with a high degree of explanatory power is
presented. Finally, the performance of five different machine
learning algorithms including two learning ensembles is compared
using various feature sets and two different datasets. The best
performances reach an F1-measure of 0.95 for music texts and
0.968 for music event texts, respectively.
Keywords–Event Detection; Text Classification; Named Entity
Recognition; Feature Selection; Ant Colony Optimization.
I. INTRODUCTION
In a highly connected world, in order to assure the rights of
artists, it is of utmost importance to develop an automatized
solution to retrospectively detect violations against copyrights
and exploitation rights related to music events. In [1], a first
approach towards such a system for online data was proposed.
This paper deepens the discussion and proposes an additional
feature selection method based on ant colony optimization
(ACO).
Individuals, groups and organizations can infringe artists’
copyrights in different ways. Whereas individuals might create
unauthorized copies, groups and organizations responsible for
public events or artists playing at them might play music
or show movies without respecting the copyright interests of
artists, either deliberately or due to ignorance of the law.
Pursuing individual interests of artists is difficult to realize
due to practical reasons (e.g., lack of information). Therefore,
authorities or private institutions are entrusted with the artists’
interests. Usually, organizers register music or movie titles,
which are going to be played at an event together with the
expected number of participants with these representatives
and by paying for the licenses receive the right to play
these titles. The official authorities and institutions of the
artists then transfer the license fees to the respective artist.
One of the largest private institutions in Germany is the
Gesellschaft f¨
ur musikalische Auff¨
uhrungs- und mechanische
Vervielf¨
altigungsrechte (GEMA, English: Society for musical
performing and mechanical reproduction rights) representing
the rights of about 2 Million artists all over the world and
with a total revenue of 1 Billion Euros a year [2]. So far,
finding unregistered events after they have taken place is very
difficult and is a process mostly done manually.
Nowadays, the information that an event is taking place is
often spread using online newspapers, Facebook, Twitter as
well as websites. Additionally, after an event has taken place
it is often discussed using the same means of communication.
Spreading the information this way is often the first choice, as
many people can be reached in a short amount of time. Hence,
analyzing these textual data makes it possible to automatically
find the information needed to uphold the artists’ rights. Text
Mining, also referred to as Text Analysis, focuses on the
analysis of texts in order to receive high level information
and latent patterns. It can be applied to many different areas
[3], however, it plays a special role in forensics and predictive
policing, where it can be used to detect events with a potential
to escalate [4]. Event detection is a specific Text Mining
problem in which texts are analyzed in order to mine a set
of texts that have a semantic link or share conceptual patterns
regarding past or future events [5] [6].
The study specifically addresses the detection of music
events announced or talked about in social media and online
newspapers, with the intention to find those events where
copyrights might be violated. Because of the vast amount of
data that needs to be taken into account, the data can only
be effectively analyzed using machine learning techniques and
methods applied in automatized text classification [7]. Since
the selection of features is a critical step due to the “curse
of dimensionality” problem [8], the Ant Colony Optimization
(ACO) algorithm was used for feature selection. To the best
of our knowledge, there have been no studies, so far, using
ACO for event detection problems. As shown in Section V-B,
when using ACO for the selection of features, the experimental
results slightly improve compared to the results achieved using
the approach proposed in [1]. Additionally, the number of
features representing the documents decrease dramatically.
This paper is organized as follows: In Section II, some re-

lated work is briefly reviewed. Section III describes difficulties
in the current domain as well as the development of a gold
standard. The proposed concept is explained in Section IV.
Details about the experimental evaluation, the results as well
as a short discussion can be found in Section V. Finally, in
Section VI a short conclusion is given and some aspects of
future work are discussed.
II. RE LATE D WOR K
As mentioned above, event detection is a special text mining
problem and can also be seen as a classification problem [9].
However, first it needs to be clarified how an event is defined.
In this study, we chose a definition based on the ontology
by [10]. Accordingly, an event is defined by the presence of
agents at a specific time and place, who are engaged with or in
a common matter (product) under concomitant circumstances
(factors). Event detection was initiated by the Topic Detection
and Tracking research program [11], yet, only focuses three
of the original five tasks, namely: tracking, detection and first
story detection [12].
In the past years, several approaches have been developed
for closed and open domains. For the former manually de-
signed keyword lists can be used to detect specific events
in texts [13]. Those keyword lists work effectively, yet need
expert knowledge to define the event-specific keywords. Fur-
thermore, keyword lists are limiting the search framework,
which is why they will not work for open domains and can
only be used as an additional resource for more complex event
types, as is the case with the detection of music events. Another
example for the detection of events within a specific field is
presented by [14] and [15], both working on the detection
of economic events that might influence the market, such as
mergers. For open domains, [6] proposed a method using
machine learning techniques, like clustering and Named Entity
Recognition (NER) combined with an ontology (DBpedia) in
order to classify Tweets into eight predefined event categories.
Similar to event extraction, the recognition of events might
also be categorized as data-driven or knowledge-driven event
recognition. In [14] and [15], data-driven approaches were
used, both taking mentions of real-world occurrences into
account in order to classify their texts into different types of
economic events. However, the data-driven approaches fail to
consider semantics. In contrast, knowledge-based approaches
focus on mining patterns from data to deliver potential rules
representing expert knowledge. Depending on the domain
or the context, linguistic, lexicographic as well as human
knowledge or a combination of these is applied [16].
Much work has been done concerning event detection using
different approaches within different fields. Certainly, some of
the proposed methods, such as those presented in [13] and
[6], can be applied for the detection of music events and our
concept is based on the work by [6]. However, in the domain of
music event detection, some difficulties appear. For example,
events might be announced only using the name of an artist.
Some of these difficulties will be discussed in later sections.
One example for a study on music events working with
Twitter data is given in [17]. In their study, they identify
musical events mentioned in Twitter in order to create a list
including sets of artists and venues. The information can be
added to an already existing list, for example, a city event
calendar [17].
III. DATA PREPARATION
Since the nature of the data is very heterogeneous – different
sources like Facebook and newspapers are considered – its
analysis has inherent challenges. Below, some of them are
discussed in more detail.
A. Data Sources
At the beginning of the study, experts, during their work on
manually detecting unregistered music events, independently
and arbitrarily preselected more than 1000 music event relevant
and irrelevant texts from Facebook and online newspapers.
This dataset was then annotated as presented below and used
as a basis for our gold standard.
B. Challenges
Noisy Data:In general, texts from social media are inher-
ently characterized by noise. For example, texts often include
web addresses, telephone numbers, dates and other characters
like hashtags. Furthermore, the texts posted, for example, on
Facebook or Twitter are not well written in terms of their
grammar and orthography. The application of standard NLP
tools to correct such mistakes may lead to incorrectly written
names of musicians. As these names are crucial for this study,
important events may not be detected.
Text Length:Due to technical restrictions and their intended
usage, texts in social media are often very short. Information
is compressed as much as possible, for example, by using
emoticons or abbreviations or by completely leaving out
words. Therefore, the application of standard text analysis
methods is often difficult, especially, if the method relies
on syntactically correct structures. Considering the following
text from Facebook, the application of standard Named Entity
Recognition methods fails, because some syntactic features are
missing:
“Foo Fighters Eintritt 19. in Hamburg”
Latent Information:Taking the example from above, the
crucial information that needs to be found is – even if the
text is already classified as an event – that Foo Fighters is a
band name and, therefore, the text announces a music event.
Typically, such information is extracted by applying methods
from the field of NER as discussed in [18]. Traditionally, NER
is a subtask in the field of information extraction that focuses
on locating structured information in a text and assigning it to
predefined categories such as names of persons, organizations
and locations. However, distinguishing normal persons from
singers or normal organizations from bands is challenging

and presents one of the biggest problems in the selection of
appropriate features as no prior information is available that
indicates whether what the NER model identified is really
music-related. This can be changed by adding additional infor-
mation in the gazetter. This means, before the classification it is
already known that, i. e., Johann Sebastian Bach is a musician.
However, a much more challenging task is the identification
of entities in a text such as musicians that are unknown, for
example, a new band or DJ. Unfortunately, texts including
these entities appear more often than texts announcing events
with known entities.
Dynamic Entities:Information is always dynamic and
changes in meaning depending on the time of production. The
latent new NER-entities (e.g., musicians, bands or groups)
change over the time. An example would be the singer and
songwriter Ed Sheeran. Before he became a known musician,
he would need to have been labeled as a normal person.
However, now he needs to be labeled as a musician. This
means, which named entities are relevant changes depending
on the point of time a text was written. This triggers the
requirement to simultaneously update the knowledge base of
our system.
C. Gold Standard
Because there are no suitable training data available, it
was necessary to create a gold standard as a basis for the
training and evaluation of various classification models. As was
mentioned above, texts were collected arbitrarily, including 21
texts from online newspapers and 1,097 texts from Facebook.
These were manually annotated as music related or music
unrelated as well as event related or event unrelated. Both
decisions were made independently of each other. Due to text-
inherent vagueness, the data was independently labeled by 35
people. In order to ensure the quality of the labeled data, each
person was only allowed to work for 2 hours a day.
The final decision regarding what category a text belongs
to was made by using a majority criterion. This criterion
requires a minimum number of people to agree on a decision
in order to provide a confident classification. If the minimum
number of agreements was not achieved for a given text,
the text was considered ambiguous and removed from the
corpus. The minimum number of agreements was derived from
a binomial test under the null hypothesis that each decision
individually made by every study participant is conducted at
random. This hypothesis thus states that p+=p−= 0.5,
where p+and p−are the decision probabilities. With respect
to the null hypothesis, for every number of agreements da
probability P(d|p+)can be derived from the corresponding
binomial distribution. The minimum number of agreements
dcrit is equal to d, where the null hypothesis can be rejected
according to P(d≥dcrit|p+)< α. Here, αcorresponds to the
Bonferroni-corrected significance level of 0.05/n, with nbeing
the number of considered texts. In this study, the minimum
number of agreements dcrit was 29 for the text corpus.
As a result, the corpus consists of 19 newspaper texts
and 867 Facebook texts. 335 out of the 867 Facebook texts
and 14 out of the 19 newspaper texts are music relevant.
Table I provides some descriptive statistics. When music event
classification is considered, the number of texts that meet the
Bonferroni constraint drops to 505, whereas 251 Facebook
texts and 9 online newspaper texts are music event relevant.
Table II provides the descriptive statistics for the music event
related data. In summary, at the end, two datasets were created:
one for music relevance, including 886 texts and one for music
event relevance with 505 texts.
TABLE I. STATISTICS OF THE DATA REGARDING MUSIC DETECTION.
# texts #tot words #avg words shortest longest
newspaper 19 2,071 109 14 387
Facebook 867 85,965 99.1 1 1,238
total 886 87,965 99.3 1 1,238
TABLE II. STATISTICS OF THE DATA REGARDING MUSIC EVENT
DETECTION.
# texts #tot words #avg words shortest longest
newspaper 13 1,077 82.85 14 277
Facebook 492 59,440 120.81 1 1,238
total 505 60,517 119.84 1 1,238
In order to describe the data in the domain of music events,
we defined an XML-schema, with which our raw data can be
concisely structured in order to serve as a gold standard to
train and test models in this field. Even though this work is
focused on music event detection, the schema is constructed to
contain various types of event data, such as music, theater, or
readings. It includes, beside others, the following information:
•raw text
•source (e. g., Facebook)
•event-related ({0,1}and certainty)
•event-type-related ({0,1}and certainty)
•event location
•event-date
•persons
•different types of roles (e. g., musician, actor)
•different types of events (e.g., music, theater)
It needs to be emphasized that the relation between any
text and a specific category is described twice: binary and
with a numeric value. The binary description refers to the
classification and thus serves as a ground truth, whereas the
numeric value represents the degree of certainty. With this gold
standard the following areas may be addressed:
•classification of texts regarding different event-types
•recognition of event-related entities, i. e., roles of per-
sons, organizations and locations
Named entities are considered because they provide strong
features for the classification, as was shown in [19] and [20].
For example, if the name Eric Clapton, an English singer and
songwriter, appears in a text, this is a strong indication that
the current text is music related. Since classic NER mostly

concentrates on distinguishing between persons, locations, and
organizations, a more detailed categorization including some
kind of prior knowledge is needed. The entire dataset was
annotated and curated manually according to the schema
described so far.
IV. PROPOSED CONCEPT
The task of detecting texts concerning music events is a
typical categorization task. Categorization, as a special case of
classification, attempts to categorize a text into a predefined set
of conceptual categories using machine learning techniques.
Formally, let T=t1, ..., tmbe a set of texts to be categorized,
and C=c1, ..., cna set of categories, then the task of catego-
rization can be described as surjective mapping f:T→C,
where f(t) = c∈Cyields the correct category for t∈T. In
the field of music event detection, texts need to be assigned
to one out of two main classes: related to a music event
or not. Texts of the former class can be further categorized
into different event types, such as public concerts. This might
be of great importance as some music, e. g., religious music
or classical music concerts, are license free or public music
resources.
Currently, institutions responsible for the enforcement of
exploitation rights have to detect unannounced music events
predominantly manually and with the help of search engines.
This leads to various problems. Firstly, the manual search
is very inefficient on large-scale data. Secondly, the manual
checking process is error-prone and differs depending on the
person who judges the data. Furthermore, the current process
chain can hardly be deployed in an online mode due to its
semi-automated nature.
unlabeled
Texts
Pre-
process
Seeds Train Classify
More
Texts? End
No
NER
Yes
Start
labeled
Texts
Feature
Selection
Gazetteer
ACO
Figure 1. The proposed workflow of music event detection.
To overcome these limitations, a semi-supervised process-
chain using a bootstrapping approach, as depicted in Figure
1, is proposed. The advantage of the chosen approach is that
the training can start with very few but highly descriptive
examples in order to create a first restrictive classifier which
will be further improved in upcoming iterations until all texts
are classified or no further improvement is possible. Next, each
step is discussed in more detail.
A. Preprocessing
As mentioned in Section III-B, the texts we worked on
mostly come from the Internet. Such texts often contain typing
errors and are often written in informal language, including
dialect. This leads to even noisier data than usual in textual
texts. Besides common shallow text preprocessing, including
stopword and punctuation removal as well as stemming or
lemmatizing, there is a strong need for additional language
information. This information can be provided in the form of
a knowledge base curated by experts. For instance, preselected
terms, such as party or live music, can be used to build the
gazetteer. Additional useful information might be venues of
interest, such as clubs or caf´
es, where music events often take
place. In short, information directly related to music events
can be used as a basis of knowledge. This knowledge base
can be a simple gazetteer, as is the case in our study, or can
incorporate more complex structures, as in [21].
B. Collecting Seed Texts
The most crucial task in bootstrapping is finding seed texts
which represent the concept of the classes as well as possible.
The usage of some kind of highly descriptive key words or
phrases collected from experts in this field is one possible way
to find seed texts in a highly accurate, but, nevertheless, very
restrictive way.
C. Named Entity Recognition (NER)
As was shown in [18]–[20], named entities might be a useful
feature for text classification tasks. In a first step, named enti-
ties are identified using any NER method, as discussed in [18].
However, as was already discussed in Section III-B, the named
entities detected this way are not specific enough. Hence,
domain-specific knowledge resources like MusicBrainz, an
open music encyclopedia, and DBpedia can serve as a music
database for distinguishing recognized entities further, in order
to assign appropriate roles to them, for example, musician
to a person. The richer and up-to-date this knowledge base, the
more accurate is the classification. The entire process of music
event related Named Entity Recognition is shown in Figure 2.
The influence of using NER with a knowledge base is clearly
shown in Section V-B.
Start
N-Grams
MusicBrainz
DBpedia
Event
Recogn.
Roles End
Classic
NER
DB
Assign.
Figure 2. The proposed workflow of detecting music related named entities.
D. Feature Selection
The next step is the selection of appropriate features to
represent the text data. Feature selection is always a critical
step in text classification tasks. On the one hand, well selected

features are necessary to achieve highly accurate results. On
the other hand, they help reduce the feature space and, as a
consequence, minimize the time complexity [22].
Generally, feature selection, as a typical machine learning
task, can be distinguished in supervised, unsupervised and
semi-supervised approaches depending on whether the data is
labeled or not [23]. The approach for music event detection
used in this paper relies, at least initially, on the availability
of labeled training data. Thus, it belongs to the algorithmic
class of supervised methods. Below, three general approaches
in supervised feature selection are briefly introduced:
1) Filter Approach: Here, the explanatory power of fea-
tures is measured by using intrinsic properties of data
to select the best feature without employing predictive
models. The principal components of methods based
on this approach are the feature search and selection
criterion. The feature can be ranked according to the
score evaluated by statistical measures such as the chi-
squared test or the information gain [22]. The ranked
features can be either removed or selected from feature
sets by comparing their explanatory power with a
given threshold. Thus, the most challenging part of this
method is to select the proper feature candidates [24].
2) Wrapper Approach: Methods based on this approach
utilize a machine learning model to evaluate the ex-
planatory power of the elements in the feature set.
Hence, the selection of proper features can be con-
sidered as a search algorithm. In this case, candidate
features are found using a search strategy like heuris-
tic search, sequential forward selection, or backward
elimination [24]. The selected features are then used to
train a predictive model and evaluate the fitness of the
selected features by utilizing hold-out data. The overall
procedure is repeated until the desired quality criterion
is met. In this approach, the trained model can be seen
as a black box and its representational bias is decreased
by repeatedly applying the feature search process (i.e.,
by means of cross validation). The feature selection
method based on ACO as proposed in this paper is
inspired by this approach.
3) Embedded Approach: Here, a linear classifier such as
SVM or logistic regression is chosen in order to learn
a predictive model and, at the same time, select the
most appropriate features. An additional regularization
function is usually included in order to constrain the
learning of the coefficients of this model. The features,
whose coefficient is non-zero, are selected [23].
Ant Colony Optimization Algorithm:Heuristic algorithms
and especially those using swarm intelligence, such as ACO,
are developed to solve combinatorial optimization problems.
They are based on the natural principles of self-organization
as a mechanism to control the behavior of the individuals of
the swarm. In this sense, the ACO is inspired by observing the
behavior of ants in an ant colony and was first reported in [25].
An ant population is able to find a solution for a combinatorial
problem, which of course might be sub-optimal, by mimicking
the ants’ forage by performing randomized walks between food
resources and their colony.
This highly coordinated behavior between the individuals in
an ant colony can be transferred to solve computational tasks.
In this case, each ant iteratively tries to find a possible solution
to solve the problem taking into account its own current heuris-
tic information as well as additional information propagated by
other ants. In this way, the population as a whole is able to find
one, though maybe sub-optimal, consensus solution. Whenever
an ant finds a candidate solution, the information about the
path this ant has walked is spread throughout the population
guiding following individuals. This kind self-organization is
made possible by the chemical odoriferous redolence, called
pheromone, which is left by each individual of the colony.
During the search, each ant contributes proportionally to the
final optimal solution. The repeated communication between
individuals happens in-directional by changing their environ-
ment. Together the pheromone paths, left behind by the inter-
action between the individuals, form a pattern which is bigger
than the pattern of each individual [26]. This collaborative
behavior pattern is called stigmergy [27].
ACO based Feature Selection:In [28], ACO based feature
selection was used to select feature subsets for different
disease data sets. An artificial neural network was trained as a
predictive model in order to evaluate each feature subset. It is
also shown that the selection of feature can improve the perfor-
mance of a classifier. In a different study a variation of ACO,
AntMiner+, was used to mine rules that provide information
about the decision making process, i.e., considering a couple
of given attributes such as amount of bank deposits, duration
of the deposits, or credit histories [29]. Consequently, a rule-
based relation is extracted in order to classify the customer’s
creditworthiness [29]. Although the time complexity of ACO
is the highest in this study, it still provides plausible results
[29].
In the field of text classification, the feature space is usually
large. ACO-based feature selection works under the assump-
tion that inter-variable relations among a reduced feature subset
represents the original data in a way that the predicted results
are accurate [28]. As shown in [30], the ACO-based feature
selection outperforms genetic algorithms, information gain,
and the chi-squared test on the Reuters-21578 data set. Similar
results were achieved by [31] for the classification of web
pages using a decision tree, Na¨
ıve Bayes and k-NN each
combined with an ACO-based feature selection method. Using
the same data set as [28], [32] also effectively applied a ACO-
based feature selection method. The main difference between
the two studies are the strategies for state transitions and
the pheromone updates. Subsequently, the feature selection
approach used in this study is based on these two studies.
When applying ACO, in a first step the problem should
be represented as a graph, whereas each node in the graph
represents a single feature in the feature space [25] [28]
[29] [32]. In this study, all the tokens are stemmed by using
the same pre-processing method as introduced in Section
IV-A. Following the principle of ACO, those features selected

by the ants, which lead to better test results, get a higher
pheromone concentration and higher heuristic values and are,
consequently, selected more frequently. Hence, there is no
additional weighting for features necessary, however, a strategy
for initializing the features in feature selection is needed. Here,
we experimented with two strategies, one using a constant
value and the other Mutual Information (MI). The features
are selected and evaluated according to their desirabilities and
contributions to the fitness. The details are shown in Section
IV-D. This process is repeated until the desired explanatory
power of the resulting feature set is reached.
Algorithm 1 Proposed Feature Selection Algorithm with ACO
Require: All features
Ensure: The best feature subsets Sk
best
1: Initial colony information τ, η
2: Generate Ants A
3: i←0
4: while i6=K-Fold do
5: Prepare training/Test data in i-Fold
6: for l= 0 to L-iteration do
7: for ant k∈ A, t ∈ T do
8: Construct feature subsets Sk(t)in i-Fold
9: end for
10: if all constructions are finished then
11: Evaluate feature subsets Sk(t)in i-Fold
12: if stop condition satisfied then
13: Return the feature subsets Sk
best
14: else
15: Update colony information τ, η
16: Reset ant memory
17: Generate Ants A
18: Go to step 5
19: end if
20: else
21: Go to step 5
22: end if
23: end for
24: i←i+ 1
25: end while
Algorithmic Details:The feature selection algorithm used
in this study is shown in Algorithm 1 and explained in more
detail below:
Initializing the Colony: The pheromone level τand the
heuristic information ρshould either be initialized with a
constant value as suggested by [28] and [32] or with more
informative values like the information gain or the Pearson
Correlation Coefficient as suggested by [33]. In this study,
the initialization of the pheromone level and the heuristic
information is done once assigning a constant value and once
assigning MI.
Generating Ants: As suggested in [32], the number
of ants #ants should equal the number of features in the
feature space #features. Unfortunately, in our case, there
are at least 10,000 features. Therefore, in order to reduce
the computational complexity only #ants =#features/100
were created. Note that we only decreased the number of ants
in the ant colony, however, the features to be crawled stayed
unchanged.
Constructing the Feature Sets: Each ant has a limited
capacity to hold the features it crawled. In [32], the roulette-
wheel schema was used to decide the size of the feature subset
Tfor each ant given a hyper-parameter µ. In this study, µis
set to µ= 0.35. This factor influences how many features
an ant can take during its trail constructing process. The next
problem is how the ants pick a particular feature (i.e., a node
in feature graph) in such a way that the classification accuracy
is maximized.
The artificial ants move between features under a pseudo-
random proportional rule. In other words, a probability deci-
sion policy guides the ants’ walk through the adjacent features
in the search space. In [32], the state transition does not
consider exploration but only exploitation, while [28] uses a
random number drawn from a uniform distribution in order to
control exploration and exploitation. In order to get informa-
tive, yet still compact features without losing generalization
the same state transition policy as the one in [28] was used in
this study.
The probability of a feature iat time step tto be selected
by ant kis defined in Equation (1):
Pk
i(t) = (arg max[τi(t)]α×[ηi(t)]β,if (q < q0)
[τi(t)]α×[ηi(t)]β
Pu∈jk[τu(t)]α×[ηu(t)]β,if (u∈jk),(1)
where factor α, η controls the influence of the pheromone
and heuristic of feature iat time step ton the transition,
respectively. The random number q∈[0,1] drawn from a
uniform distribution, controls the trade-off between exploration
and exploitation. The hyper-parameter q0forms the threshold
value for this trade-off and jkis the set of possible features
that can be taken by ant kat time step t.
Checking whether all constructions are finished: the con-
struction step repeats until all ants have finished constructing
their tours.
Evaluate Feature Subsets: In this step, each feature subset
Sk(t)is used to train an arbitrary machine learning model.
Here, the Multi-Layer-Perceptron algorithm is used with the
same parameter settings as proposed in former experiments
described in [1], where it outperformed all other considered
models.
Afterwards, the features are evaluated with hold-out data.
The pheromone concentration is proportionally updated de-
pending on the fitness of the features f(Sk(t)). Additionally,
the fitness of the features also gives information about how
good a feature is and, thus, its desirability. In this study,
a 3-fold cross validation was used to evaluate the features.
Within each fold, out of all the best local feature subsets
Sl(t), calculated during Literations, here L= 10, the best
global feature subset was determined. During an iteration, if
the fitness of the selected best local features does not change
anymore, the iteration breaks.
This way, the colony does not only focuses its search for
optimal features in a fraction of the dataset, yet in the whole

training dataset, i.e., feature space [32]. The best local and
global feature subset contribute to the fitness of features as is
explained in the following step.
When a pre-defined stop criterion is reached, the best feature
subset will be returned and the procedure is terminated. Other-
wise, the colony information such as pheromone and heuristic
of features is updated with respect to their contributions.
Updating the Colony Information: Based on the fitness
of features determined during the previous feature evaluation
step, the pheromone values and the heuristic information of
each feature are updated. In this study, the same update rules
as in [32] were applied (see Equation (2)).
τi(t+ 1) = (1 −ρ)×τi(t) + 1
mi
#ant
X
k=1
∆k
i(t) + e×∆g
i(t),(2)
where ρis the evaporation parameter of the pheromones,
mithe absolute frequency of the feature iconsidering all
feature subsets and ethe elitist parameter which decays the
contribution of features occurring in the best local feature
subsets. ∆k
i(t)is defined as the fitness of those features that are
selected by the ants during their current tour and is calculated
as shown in Equation (3).
∆k
i(t) = f(Sk(t)) if i∈Sk(t)
0otherwise.(3)
∆g
i(t)is defined as the sum of the fitness of those features
that occur in the best local feature subsets. Equation (4) shows
how it is calculated.
∆g
i(t) = f(Sl(t)) if i∈Sl(t)
0otherwise.(4)
Subsequently, both results are normalized by their occurrence
frequency. The decayed pheromone of each feature is incre-
mentally updated by considering local updates from all tours
and the penalized global updates from the elitist tours. The
heuristic of each feature is updated as shown in Equations (5)
and (6).
ηi(t+ 1) = (1 −ρ)∗ηi(t)+∆ηi,(5)
where
∆ηi=1
mi
#ant
X
k=1
f(Sk
i(t)) ×1 + φ×exp −|Sk
i(t)|
#feature ,
(6)
and φweights the exponential ratio between |Sk
i(t)|, the length
of selected feature subset by kin step tin which the feature i
occurs, and the total size of the feature space. Those features
that are associated with a reduced subset and yield a better test
performance than other features deserve a higher desirability.
Reset Ant Memory: All the features selected by ants,
their pheromone concentrations as well as their desirability
are released for the next tour.
This procedure can be finished with O(ikn)time complexity
and O(kn2)space complexity, where iis the iteration, kthe
number of ants in the colony and nthe number of features. It
can be noticed that, in this study the ACO approach includes, in
addition to the pheromone level τand the heuristic information
ηother hyper-parameters, namely, µ, α, β, ρ, e, φ, q0. While
[27] discusses how the pheromone level and the heuristic in-
formation can be determined, so far, there have been no studies
reporting how the additional parameters are set. Therefore, the
optimal settings for these parameters have to be determined
empirically.
E. Training and Classification
The final step is to train a first classifier using the seed texts
and to try to assign categories to the other texts. This step is
repeated until no improvement of the classifier can be achieved
or no remaining texts are left.
F. System Complexity
In the following section, the system complexity shall be
briefly described on the basis of time and space.
Time Complexity:The time complexity of the system,
without considering the training and classification process, can
be described as shown in Equation (7),
T(n) = Tpre +Tgazetteer +Tseeds +Tner +Tfeasel
=O(2p)+2O(1) + 3Θ(lp) + O(L|S|3) + O(ikn)
(7)
where Lis the number of samples and |S|the number of labels
in the NER process. Furthermore, lis the length of the string,
pthe length of the search pattern in the string, ithe iteration
in feature selection, kthe number of ants in the ant colony,
and nthe number of features.
Space Complexity:Similarly, the space complexity can be
measured without considering the training and classification
process as shown in Equation (8),
T(n) = Tpre +Tgazetteer +Tseeds +Tner +Tfeasel
=O(2p) + O(g) + Θ(lp) + O(s+l) + 2Θ(lp)
+O(r) + O(kn2)
(8)
where gis the size of the gazetteer, rthe number of roles, s
the size of the trained NER-model, kthe number of ants in the
ant colony and nthe number of features. After analyzing the
time and space complexity, it can be shown that the system
requires intensive resources in preprocessing and in identifying
named entities with respect to time and space complexity.
Thus, the performance of our system, regarding time and space
complexity, depends on the methods that are used in these two
setups.
V. EX PE RI ME NTAL EVAL UATION
To create first baseline results, the labeled data (see Sec-
tion III-C) were categorized using the following supervised
machine learning methods: Na¨
ıve Bayes, SVM, and MLP and
two ensemble approaches: AdaBoost and RandomForest [34].
The categorization was done once with each dataset. Firstly,

the dataset with 886 texts was used and categorized as music
relevant or not. However, as the ultimate goal is a system for
the detection of music events and not just music, secondly, the
dataset with only 505 texts was categorized as music event
relevant or not.
The performance of each experimental setting was deter-
mined using micro-averaged precision (Micro P.) and recall
(Micro R.) as well as the F1 measure.
A. Setup
In this study, only two sources of texts concerning mu-
sic events are considered: Facebook as well as daily and
weekly online newspapers. The raw data were preprocessed
as described in Section IV-A. Furthermore, all numbers, for
example, telephone numbers and dates, were removed and,
therefore, not considered in the categorization. For comparison,
two different datasets for each dataset were created. The
first dataset contains word tokens that were processed with
the Porter stemmer [35], whereas for the second dataset the
algorithm proposed in [36] was used. For the detection of
named entities a Conditional Random Field approach was
applied, as proposed by [37]. As was mentioned in Section
IV-D, MusicBrainz und DBpedia were used to assign roles to
named entities and were combined in order to increase the
number of matches.
In this study, the following four representations of the texts
incorporating different features were compared:
•multinomial Bag of Words (BoW),
•Term Frequeny Inverse Document Frequency represen-
tation of BoW (TF-IDF(BoW)),
•multinomial BoW and music event related Named Enti-
ties (BoW+NE), and
•TF-IDF representation of BoW and NE (TF-
IDF(BoW+NE)).
In case of named entities only their type (role) was con-
sidered as a feature rather than the entity itself, e. g., song
writer or musician were taken as a feature instead of Eric
Clapton. Moreover, it was only possible to train the SVM
with frequency-based features. In the ensemble approaches, an
SVM (using the same setting as reported in [1]) and a decision
tree were used as basic classifiers. This is distinguished later
by svm and tree, respectively. As a criterion to measure the
quality of split, ”Gini” for the Gini impurity and ”Entropy”
for the information gain were considered, which are labeled
with gini and ent, respectively.
B. Results
Because the features were investigated by means of ACO by
using a 3-fold cross validation due to the computation time, the
same experimental settings as in [1] were used. Additionally,
for the ACO-based selection of features the following settings
for the hyper-parameter were used: α= 0.5,β= 0.8,ρ= 0.5,
q0= 0.6,e= 0.6,φ= 0.5,ω= 0.5. The original number of
features for both datasets used in this study is shown in Table
III, while the size of the selected features is shown in Table
IV. For initializing the pheromone and heuristic in ACO, a
constant value for the first experiments was used and later MI.
TABLE III. NUMBE R OF FEATU RE S OF TH E MU SIC A ND T HE MU SI C
EV ENT DATAS ET.
# raw # raw including NEs
Music Relevance 14,275 10,921
Music-Event Relevance 12,171 9,268
TABLE IV. NUMBE R OF SE LE CTE D FE ATURE S OF T HE MU SI C AND T HE
MU SIC E VE NT DATASE T.
Dataset Initialization # raw # raw including NEs
Music Relevance Constant Value 11,164 9,584
MI 9,921 8,428
Music-Event Relevance Constant Value 8,618 7,375
MI 7,852 6,485
As can be seen in Table IV, the number of features is
considerably reduced in both datasets after using ACO to select
the features. The overall results achieved with this reduced
feature set when being used in the proposed classification
pipeline is shown below:
Baseline Results: The baseline results of the music rele-
vance decisions regarding the gold standard dataset described
in Section III-C are shown in Table V whereas the results for
the categorization of music event relevance are shown in Table
VI. The results using stemming were compared with those
achieved using lemmatization and it was observed that stem-
ming lead to slightly better results. As can be seen in Table V,
the best results for the categorization of music relevance, based
on the F1-measure, were achieved using a frequency-based
representation of words and named entities (roles) and MLP.
Comparable results were achieved by an AdaBoost model
based on SVMs as the basic classifier but considering the same
feature settings. In comparison, a combination of BoW and
named entities (roles) and an MLP model achieved the best
results for the categorization of music event relevance. These
results are presented in Table VI. Furthermore, it was found
that the best performing model and feature combination (MLP
and BoW+NE) failed when the features (words) occurred in
both, the relevant and non-relevant texts, when the texts were
very short or when there were not enough features with a high
explanatory power available.
Overall, the classification results of music relevance and
music event relevance are clearly improved when named
entities are considered as features.
Considering ACO Feature Selection: As mentioned in
Section IV-D, by using MI to initialize the pheromone and
heuristic of the features, 9921 features were found using the
886 raw documents and 8428 features using 886 documents by
considering named entities in the documents. This is repeated

TABLE V. RE SU LTS OF T HE 3- FOLD CROSS VALIDATION FOR THE MUSIC
DATASET USING STEMMING.
Model Feature Micro P. Micro R. F1
Na¨
ıve Bayes
BoW 0.656 0.974 0.784
TF-IDF(BoW) 0.991 0.625 0.766
BoW+NE 0.699 0.983 0.817
TF-IDF(BoW+NE) 0.988 0.736 0.844
MLP
BoW 0.886 0.874 0.880
TF-IDF(BoW) 0.884 0.897 0.890
Bow+NE 0.943 0.903 0.922
TF-IDF(BoW+NE) 0.934 0.931 0.933
SVM TF-IDF(BoW) 0.961 0.840 0.896
TF-IDF(BoW+NE) 0.975 0.894 0.933
AdaBoost
svm TF-IDF(BoW) 0.961 0.840 0.896
svm TF-IDF(BoW+NE) 0.975 0.894 0.933
tree TF-IDF(BoW) 0.982 0.633 0.770
svm TF-IDF(BoW+NE) 0.996 0.719 0.835
RandomForest
gini TF-IDF(BoW) 0.985 0.771 0.865
gini TF-IDF(BoW+NE) 0.990 0.851 0.915
ent TF-IDF(BoW) 0.981 0.751 0.851
ent TF-IDF(BoW+NE) 0.990 0.831 0.903
TABLE VI. RESULTS O F TH E 3-FOLD CROSS VALIDATION FOR THE MUSIC
EV ENT DATAS ET U SIN G STE MM ING .
Model Feature Micro P. Micro R. F1
Na¨
ıve Bayes
BoW 0.892 0.943 0.917
TF-IDF(BoW) 0.852 0.943 0.895
BoW+NE 0.907 0.966 0.936
TF-IDF(BoW+NE) 0.873 0.966 0.917
MLP
BoW 0.922 0.901 0.912
TF-IDF(BoW) 0.905 0.909 0.907
Bow+NE 0.953 0.924 0.938
TF-IDF(BoW+NE) 0.932 0.932 0.932
SVM TF-IDF(BoW) 0.927 0.920 0.924
TF-IDF(BoW+NE) 0.946 0.928 0.937
AdaBoost
svm TF-IDF(BoW) 0.927 0.920 0.924
svm TF-IDF(BoW+NE) 0.946 0.928 0.937
tree TF-IDF(BoW) 0.965 0.734 0.834
svm TF-IDF(BoW+NE) 0.991 0.795 0.882
RandomForest
gini TF-IDF(BoW) 0.956 0.821 0.883
gini TF-IDF(BoW+NE) 0.979 0.882 0.928
ent TF-IDF(BoW) 0.948 0.829 0.884
ent TF-IDF(BoW+NE) 0.978 0.859 0.915
by constantly initializing the pheromone and heuristic of
the features. Table V shows that with dramatically reduced
features, the best F1-measure for the music relevance dataset
is improved from 0.933 (in i.e., MLP and TF-IDF (BoW+NE))
to 0.950 (i.e., SVM and TF-IDF (BoW+NE)) when initializing
the pheromone and heuristic with MI. Furthermore, all the
results are clearly improved by using MI in the initialization,
as shown in Table VII. In comparison, the best results do not
change dramatically when initializing ACO in a constant way,
as can be seen in Table VIII. However, the feature size is
clearly reduced in comparison to the original feature size in
Table III.
The results in Tables VII and IX show that with the ACO
feature selection the performance of all machine learning algo-
rithms are improved: using MI for initializing the pheromone
and heuristic, the best baseline result for the music relevance
decision is improved from 0.933 to 0.950, and for the music
event relevance decision from 0.938 to 0.968.
By analyzing the results of detecting music-event relevance
by initializing colony information with MI, the following
conclusions were drawn:
a) The classification result can be clearly improved by
reducing the feature dimension: the average number of
tokens that occur in the documents that are correctly
classified after feature selection changes from 73 to
46 and the average classification accuracy is improved
from 0.938 to 0.954.
b) The selected features are informative for the classifica-
tion as shown by the results in Table IX.
c) The false negative classified documents are still classi-
fied incorrectly after the selection of features.
It seems that the last mentioned conclusion depends on the
distribution of the features in the documents. It was noticed
that some music related features lead to unexpected results dur-
ing the classification. For example, considering the following
document:
“Marienm¨
unster am Freitag. Abtei Marien-
m¨
unster, 15 Uhr weihnachtliche Orgelmusik, Arien
und Instrumentalst¨
ucke zum Fest.”,
words like Orgelmusik (organ music) and Instrumentalst¨
ucke
(instrumentals) give a strong feeling that this document is
somehow related to a music event in public. However, as
the features only occur once in the corpus, they fail to be
considered as strong features by the model to make a decision.
TABLE VII. RESULTS O F TH E 3-FOLD CROSS VALIDATION FOR THE
MU SIC DATAS ET U SIN G ST EMM ING A ND AC O FE ATURE S EL ECT IO N WIT H
MI INITIALIZED PHEROMONE AND HEURISTIC.
Model Feature Micro P. Micro R. F1
Na¨
ıve Bayes
BoW 0.811 0.983 0.889
TF-IDF(BoW) 0.990 0.817 0.895
BoW+NE 0.830 0.994 0.905
TF-IDF(BoW+NE) 0.981 0.885 0.931
MLP
BoW 0.948 0.894 0.920
TF-IDF(BoW) 0.915 0.926 0.920
BoW+NE 0.972 0.911 0.941
TF-IDF(BoW+NE) 0.935 0.951 0.943
SVM TF-IDF(BoW) 0.969 0.897 0.932
TF-IDF(BoW+NE) 0.967 0.934 0.950
AdaBoost
svm TF-IDF(BoW) 0.969 0.897 0.932
svm TF-IDF(BoW+NE) 0.967 0.934 0.950
tree TF-IDF(BoW) 0.987 0.645 0.780
svm TF-IDF(BoW+NE) 0.996 0.722 0.837
RandomForest
gini TF-IDF(BoW) 0.993 0.777 0.871
gini TF-IDF(BoW+NE) 0.991 0.860 0.923
ent TF-IDF(BoW) 0.983 0.762 0.859
ent TF-IDF(BoW+NE) 0.990 0.862 0.922
VI. CONCLUSION AND FUTURE WORK
In this paper, two gold standard datasets for music event
detection were presented and made publicly available here

TABLE VIII. RESULTS O F TH E 3-FOLD CROSS VALIDATION FOR THE
MU SIC DATAS ET U SIN G ST EMM ING A ND AC O FE ATURE S EL ECT IO N,
IN ITI AL IZE D WIT H A CO NSTA NT VALU E FO R PHE ROM ONE A ND H EUR IS TIC .
Model Feature Micro P. Micro R. F1
Na¨
ıve Bayes
BoW 0.667 0.974 0.792
TF-IDF(BoW) 0.979 0.673 0.798
BoW+NE 0.684 0.986 0.808
TF-IDF(Bow+NE) 0.989 0.765 0.863
MLP
BoW 0.915 0.862 0.888
TF-IDF(BoW) 0.904 0.891 0.899
Bow+NE 0.946 0.903 0.924
TF-IDF(Bow+NE) 0.921 0.940 0.930
SVM TF-IDF(BoW) 0.964 0.840 0.897
TF-IDF(Bow+NE) 0.981 0.885 0.931
TABLE IX. RESULTS O F TH E 3-FOLD CROSS VALIDATION FOR THE MUSIC
EV ENT DATAS ET U SIN G STE MM ING A ND AC O FE ATURE S EL ECT IO N WIT H
MI INITIALIZED PHEROMONE AND HEURISTIC.
Model Feature Micro P. Micro R. F1
Na¨
ıve Bayes
BoW 0.945 0.985 0.965
TF-IDF(BoW) 0.938 0.981 0.959
BoW+NE 0.949 0.985 0.966
TF-IDF(Bow+NE) 0.932 0.989 0.959
MLP
BoW 0.956 0.909 0.932
TF-IDF(BoW) 0.957 0.939 0.948
Bow+NE 0.965 0.943 0.954
TF-IDF(Bow+NE) 0.954 0.951 0.952
SVM TF-IDF(BoW) 0.966 0.970 0.968
TF-IDF(Bow+NE) 0.966 0.958 0.962
AdaBoost
svm TF-IDF(BoW) 0.966 0.970 0.968
svm TF-IDF(Bow+NE) 0.966 0.958 0.962
tree TF-IDF(BoW) 0.973 0.696 0.812
svm TF-IDF(Bow+NE) 0.986 0.795 0.880
RandomForest
gini TF-IDF(BoW) 0.965 0.844 0.901
gini TF-IDF(Bow+NE) 0.975 0.882 0.926
ent TF-IDF(BoW) 0.978 0.837 0.902
ent TF-IDF(Bow+NE) 0.975 0.894 0.933
TABLE X. RE SU LTS OF T HE 3- FOLD CROSS VALIDATION FOR THE MUSIC
EV ENT DATAS ET U SIN G STE MM ING A ND AC O FE ATURE S EL ECT IO N,
IN ITI AL IZE D WIT H A CO NSTA NT VALU E FO R PHE ROM ONE A ND H EUR IS TIC .
Model Feature Micro P. Micro R. F1
Na¨
ıve Bayes
BoW 0.882 0.935 0.908
TF-IDF(BoW) 0.865 0.947 0.904
BoW+NE 0.904 0.966 0.934
TF-IDF(Bow+NE) 0.874 0.973 0.921
MLP
BoW 0.912 0.905 0.908
TF-IDF(BoW) 0.882 0.935 0.908
Bow+NE 0.961 0.932 0.946
TF-IDF(Bow+NE) 0.940 0.947 0.943
SVM TF-IDF(BoW) 0.912 0.901 0.906
TF-IDF(Bow+NE) 0.976 0.935 0.955
[38]. Furthermore, a process chain for the categorization of
music event related texts was proposed and a first baseline
evaluation conducted. For finding much more representative
features for event detection, a couple of seed words are given
for learning extended features correspondingly. The results
change slightly in comparison to results by using the same
working process. Further experiments with ACO-based feature
selection show that a frequency-based approach considering
music specific named entities performs best together with an
SVM model for the classification of music relevant texts and
an SVM-based ensemble model for the classification of music
event relevant texts. As a strategy for initializing the relevant
parameters for ACO (e.g., heuristic and pheromone), using
MI gives promising results. As discussed in Section V-B, the
documents, in which the occurrence of non-domain related
features is more dominant than the one of domain related
features, are intended to be classified as non-music and event
relevant, although these documents are truly music and event
relevant. For this purpose, the weights of the features should
be considered in classification.
Obviously, the proposed event detection approach can also
be applied to similar domains such as movie showings, how-
ever, it should be analyzed how it may be applicable to
other more general types of events such as social events
spread via a social network, or economic events in market and
event copyright violation issues. In future work, the following
aspects need to be given more attention:
Gold standard datasets: The datasets used in this study
are relatively small, especially the one including music event
related texts and need to be extended in the future. Alterna-
tively, classification results may be improved without extended
datasets by considering further strategies. For example, transfer
learning enables the model to use pre-trained knowledge to
transfer it to the original problem domain, where there is
no sufficient training data available [39]. Based on neural
probabilistic language models [40], the texts can be represented
by using pre-trained vectors of words that enable the models
to observe the semantic in the sentences. Furthermore, active
learning shows the ability to reduce the training samples as
reported in [41].
Named Entities as Features: In [20], the authors used
named entities to represent the documents in a boolean model
and a vector space model. Similar work was conducted in [19],
where the entity power coefficient is used to create occurrences
of all terms that related to a named entity. For the study at
hand it would be important to distinguish the different named
entities into hierarchical classes, since some music events are
related to music that is not protected by any copyright laws.
In such cases, there is a need to distinguish these events
from those that might include copyright infringements. For this
purpose, a more fine-grained categorization to separate differ-
ent types of events can be realized by applying hierarchical
classification methods, such as discussed in [20] [42].
Hyper-parameter of ACO: As discussed in Section IV-D,
the ACO approach is accompanied with hyper-parameters and
different strategies for initializing the pheromone and heuristic
values. In this study, it was shown that the best results on

the music relevant dataset is clearly improved by the correct
initialization of the pheromone and heuristic values in an
ant colony. Furthermore, it would be interesting to analyze,
how the other parameter influence the results, e.g. the trade-
off between exploitation and exploration under the control
of the hyper-parameter, in this study q0= 0.6; or how the
hyper-parameter µchanges the final results, if each ant has a
bigger capacity for the feature subset, in this study µ= 0.35.
Some conclusions about how to select hyper-paremeters like
β, ρ for other applications, e.g., for the well-known traveling
salesman problem are drawn by [27]. Additionally, the optimal
population size of ants in a colony in the feature selection
process was discussed in [30]. How the other parameters are
initialized in selecting the features, should also be investigated
in the future.
Knowledge Base: There are music-related named entities
that occur in German texts but there is no corresponding
entry in databases like dbpedia or Musicbrainz for those
entities. Hence, the availability of music-related entities in such
databases plays a crucial role in expanding the knowledge base
in order to assign correct music roles. Here, suitable fall-back
strategies should be considered in future work.
Polysemy & Synonymy: During the experiments it has
been noticed that no text representation captures phenomena
like polysemy or synonymy of words. In German, the word
spielen (play) has the following syntactic and semantic envi-
ronment: “Orgel spielen” (playing organ music) and “gegen
eine Mannschaft spielen” (playing against a team). For the
former one, it is certain that it is music relevant, the latter
implies a sport event. However, without further contextual
information, even the first one might not be recognized as
music relevant. Similarly, the problem of synonymy deserves
much more attention in future work, e.g., “... in der Gosecker
Schloss-Krypta klangen wieder kristallene Engel...” (in the
Gosecker castle-crypt crystal angels sounded again). In this
context, the word klangen means to play music. Those words
are quite often used in public announcement in Germany. Thus,
it could be a meaningful feature to decide whether a text is
music relevant or not.
ACKNOWLEDGMENT
The authors would like to thank the deecoob GmbH for
acting as experts during the creation of the gold standard and
providing data. The project was funded by the German federal
ministry for economics and energy.
REFERENCES
[1] J. Xi, M. Spranger, H. Siewerts, and D. Labudde, “Towards an au-
tomated system for music event detection,” in SEMAPRO 2018, The
Twelfth International Conference on Advances in Semantic Processing.
ThinkMind, 2018, pp. 22–27.
[2] GEMA, “Gesch¨
aftsbericht mit Transparenzbericht 2017,”
https://www.gema.de/fileadmin/user\upload/Gema/geschaeftsberichte/
GEMA\Geschaeftsbericht2017.pdf, 2018 [retrieved: September,
2018].
[3] V. Gupta and G. S. Lehal, “A Survey of Text Mining Techniques
and Applications,” Journal of Emerging Technologies in Web
Intelligence, vol. 1, no. 1, 2009, pp. 60–76. [Online]. Available:
www.alerts.yahoo.com
[4] M. Spranger, H. Siewerts, J. Hampl, F. Heinke, and D. Labudde, “SoNA:
A Knowledge-based Social Network Analysis Framework for Predictive
Policing,” International Journal On Advances in Intelligent Systems,
vol. 10, no. 3 & 4, 2017, pp. 147–156.
[5] K. Giridhar and A. James, “Text Classification and Named Entities for
New Event Detection,” in Proceedings of the 27th Annual International
ACM SIGIR Conference on Research and Development in Information
Retrieval. ACM, 2004, pp. 297–304.
[6] A. Edouard, “Event detection and analysis on short text messages,”
Ph.D. dissertation, Universit´
e Cˆ
ote D’Azur, 2017, https://hal.inria.fr/tel-
01680769/document [retrieved: September, 2018].
[7] M. Ikonomakis, S. Kotsiantis, and V. Tampakas, “Text Classification
Using Machine Learning Techniques,” WSEAS Transactions on Com-
puters, vol. 4, 2006, pp. 966–974.
[8] R. Bellman, “The Theory of Dynamic Programming,” Bulletin of the
American Mathematical Society, vol. 60, no. 6, 1954, pp. 503–515.
[9] Y. Yang, T. Pierce, and J. Carbonell, “A Study of Retrospective and On-
line Event Detection,” in Proceedings of the 21st Annual International
ACM SIGIR Conference on Research and Development in Information
Retrieval. ACM, 1998, pp. 28–36.
[10] Y. Raimond and S. Abdallah, “The Event Ontology,” http://motools.
sourceforge.net/event/event.html, 2007 [retrieved: July, 2019].
[11] J. Allan, Topic detection and tracking: event-based information organi-
zation, J. Allan, Ed., 2002.
[12] M. Cordeiro and J. Gama, “Online social networks event detection:
A survey,” in Solving Large Scale Learning Tasks. Challenges and
Algorithms, S. Michaelis, N. Piatkowski, and M. Stolpe, Eds., vol. 9580.
Springer Verlag, 2016, pp. 1–41.
[13] T. Sakaki, M. Okazaki, and Y. Matsuo, “Earthquake shakes twitter users:
Real-time event detection by social sensors,” in Proceedings of the 19th
International Conference on World Wide Web. ACM, 2010, pp. 851–
860.
[14] E. Lefever and V. Hoste, “A classification-based approach to economic
event detection in dutch news text,” in Proceedings of the Tenth Interna-
tional Conference on Language Resources and Evaluation. European
Language Resources Association (ELRA), 2016, pp. 330–335.
[15] G. Jacobs, E. Lefever, and V. Hoste, “Economic event detection in
company-specific news text,” in Proceedings of the First Workshop
on Economics and Natural Language Processing. Association for
Computational Linguistics, 2018, pp. 1–10.
[16] F. Hogenboom, F. Frasincar, U. Kaymak, and F. de Jong, “An overview
of event extraction from text,” in CEUR Workshop Proceedings, vol.
779, 2011, pp. 48–57.
[17] E. Benson, A. Haghighi, and R. Barzilay, “Event discovery in social
media feeds,” in Proceedings of the 49th Annual Meeting of the Asso-
ciation for Computational Linguistics: Human Language Technologies
- Volume 1, ser. HLT ’11. Stroudsburg, PA, USA: Association for
Computational Linguistics, 2011, pp. 389–398.
[18] D. Nadeau and S. Sekine, “A survey of named entity recognition and
classification,” Linguisticae Investigationes, vol. 30, 2007, pp. 3–26.
[19] M. K. Stefan Andelic, “Text classification based on named entities,”
2017, pp. 23–28.
[20] Y. Gui, Z. Gao, R. Li, and X. Yang, “Hierarchical text classification
for news articles based-on named entities,” Advanced Data Mining and
Applications, vol. 7713, 2012, pp. 318–329.
[21] M. Spranger and D. Labudde, “Towards Establishing an Expert System
for Forensic Text Analysis,” International Journal on Advances in
Intelligent Systems, vol. 7, no. 1/2, 2014, pp. 247–256.
[22] Y. Yang and J. O. Pedersen, “A comparative study on feature selection
in text categorization,” in Proceedings of the Fourteenth International
Conference on Machine Learning, 1997, pp. 412–420.

[23] J. Tang, S. Alelyani, and H. Liu, “Feature selection for classification: A
review,” Data classification: Algorithms and applications, 2014, p. 37.
[24] H. Liu and H. Motoda. Boca Raton, Florida: Chapman and Hall/CRC,
Oktober 2007, ch. 2,13.
[25] M. Dorigo, “Ant Colony System: A Cooperative Learning Approach to
the Traveling Salesman Problem,” IEEE Transactions on Evolutionary
Computation, vol. 1, April 1997, pp. 53–66.
[26] D. J. T. Sumpter, Collective Animal Behavior. Princeton (New Jersey):
Princeton University Press, 2010, ch. 1,5.
[27] M. Dorigo and T. St¨
utzle, Ant Colony Optimization. Cambridge,
Massachusetts. London, England: The MIT Press, 2004, ch. 1,3.
[28] R. K. Sivagaminathan and S. Ramakrishnan, “A hybrid approach for
feature subset selection using neural networks and ant colony optimiza-
tion,” Expert Syst. Appl., vol. 33, 2007, pp. 49–60.
[29] D. Martens, M. De Backer, R. Haesen, J. Vanthienen, M. Snoeck,
and B. Baesens, “Classification With Ant Colony Optimization,” IEEE
Transactions on Evolutionary Computation, vol. 11, no. 5, 2007, pp.
651–665.
[30] M. H. Aghdam, N. Ghasem-Aghaee, and M. E. Basiri, “Text feature
selection using ant colony optimization,” Expert Systems with Appli-
cations, vol. 36, no. 3, April 2009, pp. 6843–6853.
[31] E. Sarac¸ and S. A. ¨
Ozel, “An Ant Colony Optimization Based Feature
Selection for Web Page Classification,” The Scientific World Journal,
vol. 2014, 2014.
[32] M. M. Kabir, M. Shahjahan, and K. Murase, “An Efficient Feature Se-
lection Using Ant Colony Optimization Algorithm,” Neural Information
Processing, vol. 36, 2009, pp. 242–252.
[33] M. M. Kabir and K. Murase, “A Wew Wrapper Feature Selection
Approach Ssing Neural Network,” Neurocomputing, vol. 36, 2010, pp.
3273–3283.
[34] F. Pedregosa et al., “Scikit-learn: Machine learning in python,” J.
Mach. Learn. Res., vol. 12, Nov. 2011, pp. 2825–2830. [Online].
Available: http://dl.acm.org/citation.cfm?id=1953048.2078195
[35] M. F. Porter, “An algorithm for suffix stripping,” Program, vol. 14,
no. 3, 1980, pp. 130–137.
[36] H. Schmid and H. Schmid, “Improvements in part-of-speech tagging
with an application to german,” Proceedings of the ACL SIGDAT-
Workshop, 1995, pp. 47–50.
[37] J. R. Finkel, T. Grenager, and C. Manning, “Incorporating non-local
information into information extraction systems by gibbs sampling,” in
Proceedings of the 43rd Annual Meeting on Association for Computa-
tional Linguistics. ACM, 2005, pp. 363–370.
[38] J. Xi, “Music Classification Gold Standard Datasets,” https://github.
com/fosil-mw/music classification data, 2018 [retrieved: July, 2019].
[39] Q. Yang and S. J. Pan, “A Survey on Transfer Learning,” IEEE
Transactions on Knowledge Data Engineering, vol. 22, 2009, pp. 1345–
1359.
[40] Y. Bengio, R. Ducharme, P. Vincent, and C. Janvin, “A neural proba-
bilistic language model,” Journal of Machine Learning Research, vol. 3,
2003, pp. 1137–1155.
[41] D. K. Xiao Li and C. X. Ling, “Active learning for hierarchical text
classification,” Tan PN., Chawla S., Ho C.K., Bailey J. (eds) Advances
in Knowledge Discovery and Data Mining, vol. 7301, 2012, pp. 14–25.
[42] A. McCallum, R. Rosenfeld, T. M. Mitchell, and A. Y. Ng, “Improving
text classification by shrinkage in a hierarchy of classes,” in Proceedings
of the Fifteenth International Conference on Machine Learning, 1997,
pp. 359–367.