Download 4. support vector machines

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Data assimilation wikipedia , lookup

Discrete choice wikipedia , lookup

Time series wikipedia , lookup

Choice modelling wikipedia , lookup

Regression toward the mean wikipedia , lookup

Coefficient of determination wikipedia , lookup

Regression analysis wikipedia , lookup

Linear regression wikipedia , lookup

Transcript
A Comparative Study of the various methods used for the
Prediction of Hit Songs
Aishwarya Harne
Mihir Borkar
Dept. of Computer Engineering
D.J. Sanghvi College of
Engineering
Mumbai University, Mumbai
Dept. of Computer Engineering
D.J. Sanghvi College of
Engineering
Mumbai University, Mumbai
[email protected]
[email protected]
Abhinav Garg
Prof. Abhijit Patil
Dept. of Computer Engineering
D.J. Sanghvi College of
Engineering
Mumbai University, Mumbai
Dept. of Computer Engineering
D.J. Sanghvi College of
Engineering
Mumbai University, Mumbai
[email protected]
[email protected]
ABSTRACT
The music industry today is filled with scores of struggling
composers and music producers churning out a plethora of
songs. Hardly 5% of these tracks are actually released and
even fewer become popular amongst the masses. This project
aims to support the 20 billion-dollar music industry cater to
those who patronize it by helping musicians understand the
complete extent of how impactful their new soundtrack will
be in terms of commercial popularity. We propose to do this
by using logistic regression to determine the popular features
of a track encoded in a numerical format. We do this by
essentially extracting the frequency, tempo and pitch of the
sound from the MIDI file. The MIDO parser in Python
encodes these characteristics in a numerical format. The
MIDO file associates every instrument with a number (1 for
Drums, 2 for Guitar etc.) separating consecutive notes with
colon ( : ) as a delimiter. We proceed to increase the accuracy
of the training by using 50 soundtracks to create positive
learning and 50 to create a negative one.
three models, namely Logistic Regression, Naive Bayes and
the SVM (Support Vector Machine). All the three methods
help us work on a dichotomous result. We proceed to look at
each of these methods in depth and then finally conclude on
which method is most applicable to the project.
2. TYPE OF MUSICAL FEATURES
The characteristics of a particular track can be identified by
three basic components that which make it unique when
compared to other musical tracks as well as increase its
popularity and that are [2] :
1.
Timbre/instrument
It is also known as tone color or tone quality in
psychoacoustics. Timbre is the quality of the musical
note, sounds that differentiates the various types of
sounds production which include string instruments,
voices and musical instruments, wind instruments and
percussions.
2.
Melody
It is a sequence of musical notes that is played together
as a part of a composition. The arrangement of single
notes forms an important aspect of the Melody.
3.
Beats
It is a periodic variation of amplitude or sound. It is a
regular rhythmic sound or movement that forms the
backbone of any composition.
Keywords
Naïve Bayes, logistic regression, security protocols, support
vector machine
1. INTRODUCTION
In this paper, we seek to establish the best way to determine
the commercial success of a song. After decoding the MIDI
files into a numerical format, there are a variety of
probabilistic models that can be applied to analyze and
determine the success of the sound track. We are considering
3. LOGISTIC CLASSIFICATION
A logistic regression model can be used in classification of
inputs into binary or multiple categories. Logistic regression
measures the relationship between the categorical dependent
variable and one or more independent variables by estimating
probabilities using a logistic function, which is the cumulative
logistic distribution. Thus, it treats the same set of problems as
probit regression using similar techniques, with the latter
using a cumulative normal distribution curve instead.
Equivalently, in the latent variable interpretations of these two
methods, the first assumes a standard logistic distribution of
errors and the second a standard normal distribution of errors.
Logistic regression can be seen as a special case of
generalized linear model and thus analogous to linear
regression. The model of logistic regression, however, is
based on quite different assumptions (about the relationship
between dependent and independent variables) from those of
linear regression. In particular the key differences of these two
models can be seen in the following two features of logistic
regression.
First, the conditional distribution
is a Bernoulli
distribution rather than a Gaussian distribution, because the
dependent variable is binary. Second, the predicted values are
probabilities and are therefore restricted to (0, 1) through the
logistic distribution function because logistic regression
predicts the probability of particular outcomes [1].
it a non-probabilistic binary linear classifier.In addition to
performing linear classification, SVMs can efficiently
perform a non-linear classification using what is called the
kernel trick, implicitly mapping their inputs into highdimensional feature spaces. [6]
SVMs belong to the family of linear classifiers. A linear
classifier is used to achieve the results of statistical
classification to use an object’s characteristics and identify the
group or class it belongs to. This decision is achieved based
on the value of linear combination of the characteristics.
The SVM algorithm has been widely used in text and
hypertext classification. Also, hand written characters can be
efficiently recognized by a SVM model. SVMs are also useful
in medical science to classify proteins with up to 90% of the
compounds classified correctly.
5. NAIVE-BAYES
5.1 Concept of Conditional Probability:
Logistic regression can be used to predict songs by classifying
them as hit or not by using a decision boundary which will be
based on the labels like whether the song has been constantly
rated in the top 40 of the billboard charts or has been rated in
top 100 of music websites such as Spotify.
Conditional Probability is a concept that determines the
likelihood of a particular event occurring given that another
event has already occurred.
The following formulae are used while performing logistic
classification:
Examples of this type can be solved using conditional
probability.
For E.g. - In a box that consists of apples and oranges, some
of the fruits are corrupted. What is the probability that a fruit
picked at random is corrupted if it is an apple?
5.2 Bayes Rule
Bayes rule gives quantifies the conditional probability of an
event by giving a relation involving its reverse form. [7]
Taking the above example in mind,
J (Ɵ) stands for the learning rate
P (A) = Probability that the fruit is corrupted
hƟ(x) stands for hypotheses
P (B) = Probability that the fruit is an apple
P (B|A) = Probability that the fruit is an apple, given its
corrupted
4. SUPPORT VECTOR MACHINES
In machine learning, support vector machines (SVMs, also
support vector networks[1]) are supervised learning models
with associated learning algorithms that analyze data and
recognize patterns, used for classification and regression
analysis. an SVM training algorithm builds a model that
assigns new examples into one category or the other, making
Let’s assume
P (A) = 0.3
P (B) = 0.5
P (B|A) = 0.2
According to Bayes theorem,
P (A|B) = P (B|A)*P (A)
---------------P (B)
= 0.2*0.3
--------0.5
= 0.1
5.3 NAIVE BAYES
Naive Bayes:
The Naive Bayes is generally used when trying to predict an
outcome in the presence of more than one evidence.
We decompose the Bayes rule in such a scenario and apply it
to every evidence individually.
Posterior Probability = Prior Probability * Likelihood
-----------------------------------Evidence
6. CONCLUSION
We have tried to train our data sets using Logistic regression,
SVM and Naive Bayes method. However, we conclude the
Logistic Regression to be the most effective method in this
case.
SVM is generally used for cases where time efficiency is
critical [2], but is not needed when working on a relatively
small dataset.
Naive Bayes is effective when one feature does not depend on
another [2], but in the case of music tracks, the popularity is
determined by the succession of notes. Hence this method is
not very useful.
Logistic Regression is a perfect choice since it also allows us
to impose a stricter criteria on popular songs by choosing an
appropriate probability coefficient [2].
REFERENCES
[1] Logistic Regression for Classification, available at:
https://en.wikipedia.org/wiki/Logistic_regression
[2] Wang, Keven Kedao. "Predicting Hit Songs with MIDI
Musical Features.", cs229.stanford.edu
[3] Koenigstein, Noam, Yuval Shavitt, and Noa Zilberman.
"Predicting billboard success using data-mining in p2p
networks." Multimedia, 2009. ISM'09. 11th IEEE
International Symposium on. IEEE, 2009.
[4] Shapiro, Heather. "A Bayesian
Understanding Music Popularity." (2015).
Approach
to
[5] Dhanaraj, Ruth, and Beth Logan. "Automatic Prediction of
Hit Songs." ISMIR. 2005.
[6] Support Vector Machines, available at:
https://en.wikipedia.org/wiki/Support_vector_machine
[7] Naïve Bayes Classifier, available at:
https://en.wikipedia.org/wiki/Naive_Bayes_classifier