APPLICATION OF AN ANT COLONY ALGORITHM

For Song Categorising using Metadata

Nadia Lachetar

and Halima Bahi

Computer Science Department, Skikda University, Skikda, Algeria

Labged Laboratory, Computer Science Department, Annaba University, Annaba, Algeria

Keywords: Audio indexing, Naive Bayes algorithm, Ant colony algorithm, Song categorisation.

Abstract: Instead of the expansion of the information retrieval systems, the music information retrieval domain is still

an open one. One of the promising areas in this context is the audio indexing databases. This paper

addresses the problem of indexing database containing songs to enable their effective exploitation. Since,

we are interested with songs databases, it is necessary to exploit the specific structure of the song in with

each part plays a specific role. We propose to use the title and the artist particularities (in fact each artist

tends to compose or sing a specific genre of music). In this article, we present our experiments in automated

song categorisation, where we suggest the use of an ant colony algorithm. A naive Bayes algorithm is used

as a baseline in our tests.

1 INTRODUCTION

Nowadays music is playing an important role in

human’s life, whereas digital catalogues rapidly

become larger and more inconvenient and inefficient

to access. If we do not have a good method to

explore music, a large amount of music will be

fallen into oblivion (Dang and Shirai,

2009). As the

multimedia content is growing, and digital music

libraries are expanding at an exponential rate, it is

important to secure effective information access and

retrieval.

In this paper, we propose to construct an

automatic system of categorisation of songs by

theme using metadata.

Currently many studies are devoted to the use of

acoustic information to detect the theme of songs (Li

and Ogihana, 2004; Chua and Lu, 2004; Liu and

Zang, 2003). To index an audio document, the first

step is to determine the type of information present.

In the case of songs, many studies have been

performed to detect the music, speech, or sound

features (Scheirer and Slaney, 1997; Karjalainen and

Tolonen, 1999). Very little has been done on the

song (Arroabarrem et al., 2003).

Many aspects of the music itself (such as lyrics,

genre, key or era) are shared on the Internet. Digital

music can hold information such as artist, track

name, year, and album in the source.

Our study is based on the use of metadata to

detect a theme. We propose to use the song title that

briefly describes its theme but also the

characteristics of the artist, because every artist has a

tendency to compose or sing a particular kind of

music. We introduce methods for supervised

learning classification of songs based on metadata

and we propose the use of an ant colony algorithm

for classification of songs using the title and the

characteristics of the artist.

In section 2, we describe the state of the art.

Then in section 3 we present our training data.

Section 4 is devoted to categorize songs while

section 5 presents the naive Bayes algorithm. In

Section 6 we present our approach for indexing

songs using the songs title and artist features.

Section 7 presents the obtained results and a

discussion.

2 STATE OF THE ART

Many works are devoted to the extraction of features

of a song for the descriptions of its contents are

generally guided by the acoustic analysis (Li and

Ogihana, 2004; Chua and Lu, 2004; Liu and Zang,

2003). Knees and others have a pioneering work to

build an automatic search that is able to find the

music that satisfies arbitrary queries in natural

379

Lachetar N. and Bahi H..

APPLICATION OF AN ANT COLONY ALGORITHM - For Song Categorising using Metadata.

DOI: 10.5220/0003631303710376

In Proceedings of the International Conference on Knowledge Discovery and Information Retrieval (KDIR-2011), pages 371-376

ISBN: 978-989-8425-79-9

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

language

(Knees et al., 2000). Another work

described in (Harb and Chen, 2003) based on an

automatic segmentation of the soundtrack music or

speech, using a technique of segmentation into

sentences. The music segments are indexed in a way

that allows a search by similarity. Other jobs using a

classification according to the mood of the songs are

described in (Dang, Shirai, 2009, Kanters, 2009,

Laurier and al., 2008). The classification according

to the mood does not seem interesting to apply it to a

search engine for music because the mood is

metadata subjective words are short and contain

many metaphors that can be understood by humans.

Through this work, we introduce a new

dimension of classification, considering contextual

information about the artist. Thus, each artist sings

songs with a specific emotion, such as Eric Clapton

sings sad songs often but Bob Marley likes to sing

happy songs.

3 CONSTRUCTION

OF TRAINING DATA

In this section we describe how to prepare our

training data, the collection of songs tagged with the

theme described by the title and artist features.

A great blog site Live Journal (www.livejournal.

com) is used, each blog entry is labeled with the

theme of the song given by the title of this latter.

The song title and artist features can be obtained by

simple string matching with the database artist,

obtained from open artist got from the music site

(www.musicmoz.org). The lyrics may be obtained

from the Site (www.lyrics.com).

4 SONG CATEGORIZATION

Research in the field of automatic categorization

remains relevant today since the results are still

subject to improvements. For some tasks, the

automatic classifiers perform almost as well as

humans, but for others the gap is even greater. At

first glance, the main problem is easy to grasp. On

one hand, we are dealing with a bank of songs and

on the other with a set of categories. The goal is to

make a computer application which can determine to

which category belongs a song based on its contents.

The set of categories is determined in advance. The

problem is to group the songs by their similarity.

There are two approaches to solving the problem of

songs categorization: the information using either

acoustic or verbal information. In this paper we will

focus on the words comprising the title of the song

to determine its theme and the characteristics of the

artist to determine what kind of music.

The categorization process includes the

construction of a prediction model that receives in

input the title of the song, and as output it combines

one or more labels.

Prior coding of song is necessary because there is

currently no method of learning which can directly

handle unstructured data in the model construction

stage, or when used in classification.

For most learning methods, we must convert all

texts in a PivotTable "individuals-variables".

In song categorization, we transform the title of

the song into a vector d

= d

, w

, ..., w

| T | j

where T is the set of terms (descriptors) that appear

at least once in the corpus (the collection) learning.

The weight w

correspond to the contribution of

terms t

to the semantics of title of song d

Once we choose the components of the vector

representing the song j, we must decide how to

encode each coordinate of the vector d

. There are

different methods to calculate the weight w

. These

methods are based on two observations:

More the term t

is frequently in a title of song

, more it is relevant to the subject of this song.

More often the term t

is in a collection, unless it

is used as discriminating between songs.

The Coding terms frequency x inverse document

frequency and Coding terms TFC are the most used.

5 NAIVE BAYES ALGORITHM

In machine learning, different types of classifiers

have been developed to achieve maximum degree of

precision and efficiency, each with its advantages

and disadvantages. But, they share common

characteristics (Sebastiani, 2002).

Naive Bayes Classifier is the most commonly

used algorithm. When we apply the naïve Bayes for

a song categorization task, we look for the

classification that maximizes the probability of

observing the words of titles of the songs.

During the training phase, the classifier

calculates the probability that a new song belongs to

this category based on the proportion of training

songs belonging to this category. It calculates the

probability that a given word is present in a title of

the song, knowing that this song belongs to this

category. Then as a new song should be classified,

we calculate the probability that it belongs to each

class using Bayes rule and the probabilities

KDIR 2011 - International Conference on Knowledge Discovery and Information Retrieval

380

calculated in the previous step. The likelihood to be

estimated is:

p(c

, a

, ..., a

)

(1)

Where c

is a category and a

is an attribute

Using the Bayes theorem, we obtain:

p







,



,..,=









,

,

…,



\





∗(



)

(



,



,…,



)

(2)





,



,



,…,



\



= 



\











(3)

6 APPLICATION OF AN ANT

COLONY ALGORITHM FOR

SONG CATEGORIZATION

6.1 Introduction

The originality of our approach is on adapting an

algorithm of ant colony to song categorization.

The algorithm of ant colony optimization is

inspired by the behaviour of ants searching for food.

Its principle is based on the behaviour of individual

ants; they are able to determine the shortest path

between their nest and a food source using the

pheromone which is a substance that ants lay on the

floor when they move. When an ant has to choose

between two directions, it chooses with higher

probability (Solnon, 2005).

6.2 Principe of the Algorithm

It relies on the specific behaviour of ants, and

determines the shortest path between the nest and a

food source overall progress algorithm.

6.3 For Song Categorization

For the construction of the graph, the nodes

represent titles of songs. The pheromone is a

measure of similarity between titles of songs which

may be the distance between these documents. The

choice of distance is an important parameter.

We add to the vector of features once

standardized the characteristics of the author (the

type of music and the theme he sings in general).

6.3.1 Calculate the Distance between the

Song Filed and the Songs constituting

the Graph

For our approach we use the cosine similarity

between two songs a and b defined by







(



)

∗



(



)



∑





(



)



∗

∑





()



∈∈

∈

(4)

Where:

• T is the set of attributes.

• p

(a) is the weight of term t in title of song a.

• p

(b) is the weight of term t in title of song b.

This measure allows comparing titles of songs of

different lengths by normalizing their vector.

We use the cosine similarity between each title

of song “a” of the graph of songs and the input title

of song “b” to be classified.

The following algorithm computes the cosine

similarity based on relevant attributes for the various

couples forming the nodes of a graph and the input

title of song. It takes as input the graph of songs and

the document to classify and returns as output a

similarity matrix based on relevant attributes.

Algorithm Cosine_Similarity

Begin

Input:song_Graph,song_class //graph

of songs,Classified song;

Output: Mat_Sim

// similarity matrix based on the

relevant attributes

Mat_Sim ← 0;

Begin

For each node of song_Graph

// Extract set of attribute nodes of

the graph

SIM=Calcul_Sim (node, song_class);

Mat_Sim=Mat_Sim+Sim(node,song_class)

;

Return Mat_Sim

End.

6.3.2 Ant Colony Optimization

To find the song category, we adopt the algorithm of

ant colony optimization (ACO), proposed in

(Solnon, 2005). Although the ant colony algorithm is

originally designed for the travelling salesman

problem, it finally offers great flexibility. Our choice

is motivated by the flexibility of the metaheuristics

which makes possible its application to different

problems that are common to be NP-hard. Thus the

use of a parallel model (colonies of ants) reduces the

computing time and improves the quality of

solutions for categorization.

Formalization of the problem: In our context, the

problem of classifying a song reduces the problem

of subset selection (Solnon, 2005), and we can

formalize the pair (S, f) such that:

APPLICATION OF AN ANT COLONY ALGORITHM - For Song Categorising using Metadata

381

• S contains all the cosine similarities calculated

between the graph of songs and the song to

classify. It's "matrix similarity" mat_sim.

• F is defined by the function score, the score

function is defined in (Solnon, 2005) by the

formula.



(





)



(song_Graph∩

song_class−

(





)

(5)

Splits (S') is the set of nodes in the graph which

are more similar to the song to classify. So the result

is a consistent subset S' of nodes, as the score

function is maximized.

6.3.3 Construction of the Graph of Songs

To adapt our approach, a direct graph G( V,E) is

drawn, where V is a set of vertices and E is a set of

possible edges between these vertices as shown in

figure 1 in this graph a number of ants is managed

for tmax iteration.

Figure 1: Graph representation of song categorization.

A graph G(V,E) is automatically generated from a

text file that contains the problem’s data ( words of

titles of songs).

- Graph’s vertices are bundled with words of

titles of songs .

- Graph’s edges are bundled with the initial

pheromone trails. The pheromone is a measure of

similarity between titles of songs. We use the cosine

similarity between two songs according to (4)

6.3.4 Description of the Algorithm

At each cycle of the algorithm, each ant constructs a

subset. Starting from empty subset, ants at each

iteration add a couple of nodes from the similarity

matrix. S

chosen among all couples not yet

selected. The pair of nodes to add to S

is chosen

with a probability which depends on the trail of phe-

romones and heuristics. One aims to encourage cou-

ples who have the greatest similarity and the other is

to encourage couples who are most increase the

score function. Once each ant has built its subset, a

local search procedure start to improve the quality of

the best subset found during this cycle. Pheromone

trails are subsequently updated based on the subset

improved. Ants stop their construction when all pairs

of candidate nodes are decreased the score subset.

Construction of a solution by an ant: The

following code describes the procedure followed by

ants to construct a subset. The first object is selected

randomly. The following items are selected in all

candidates.

Algorithm Construction_subset

Begin

Input:SS_problem(S,s

consistant

,f)and an

associated heuristic function:

S*P(S)→ IR

; heuristic pheromone; and

an phenomenal factor  and 2

heuristic factors 



and 



;

Numeric parameter , 



and 



Output: a consistent subset S’∈ S

Initialize pheromone trails to τmax

Repeat

For each ant k in 1 .. nbAnts,

construct a solution S

as follows:

1. Randomly select the first node O

∈

2. S

← {o

}

3. Candidat ← {o

∈S /S

∪ {oj}∈

consistent

}

4. While Candidates ≠ ∅ do

5. Choose a node o

∈ candidat with

probability

6.



∑





(



)

∗



(



)



∑





(



)



∗

∑





()



∈∈

∈

7. s

← S

union {o

}

8. Remove o

from Candidates

9. Remove from candidates each node

as S

∪{o

} ∈ S

consisting

10. End while

11. End for

Update pheromone trails according to

, ..., S

nbAnts

}

If a pheromone trail is less than

min

then set it to τ

min

Else If a pheromone trail is greater

than τ

max

then set it to τ

max

Until maximum number of cycles

reached or solution found.

End.

Nest

KDIR 2011 - International Conference on Knowledge Discovery and Information Retrieval

382

7 RESULTS AND DISCUSSION

To evaluate performances of our suggestion, we

make some experiments using two corpus one for

the training and the other for the test. We also use

the Naïve Bayes classifier as baseline one.

Table 1: Classes of corpus.

Class Nb of songs in

training stage

Nb of songs in

test stage

National anthem 12 10

Loves songs 60 46

Religious songs 36 30

Sport songs 27 26

Learning songs 18 16

The results of classification stage are reported

below for ant colony algorithm and naïve Bayes

algorithm

Table 2: Results of tests with ant colony algorithm.

Clas Nati

Ant

Lov

son

Reli

son

Spor

son

lear

son

Total

Nati Ant 9 0 1 0 0 10

Lov Son 1 42 2 0 1 46

Reli Son 1 2 26 0 1 30

Spor Son 0 1 1 23 1 26

lea son 0 1 1 0 14 16

Table 3: Results of tests with naïf Bayes algorithm.

Clas Nati

ant

Lov

son

Reli

son

Spor

son

lear

son

Total

Nati Ant 7 2 1 0 0 10

Lov Son 1 40 2 1 2 46

Reli Son 2 4 22 1 1 30

Spor Son 0 2 1 22 1 26

lea son 1 1 1 0 13 16

Precision and recall are the most used

measurements to evaluate information retrieval

systems, they are defined as follow:

Table 4: Contingency table based evaluation of the

classifiers.

Song belonging

to the category

Song not

belonging to the