view Report/chapter2/background.tex @ 23:45e6f85d0ba4

List of clips downloaded

\chapter{Background research}
Recommender systems set up opportunities and challenges for industry to understand consumption behaviour of users. In particular, for music industry, the develop of recommender systems could improve sales for artists and labels, and the discovery of new songs for listeners. However, regarding that music tastes vary from one person to another person, an advantageous music recommender system should be able to infer listeners needs through their historical listening preference information, similarities with another listeners, and audio signal features from their music collections.

In the following sections, the importance of online social networks for retrieving user-item information among with previous work on music recommender systems are presented. Subsequently, a novel approach of an hybrid recommender system based on Estimation of Distribution Algorithm (EDA) is introduced and examined.

\section{Online Social Networks}
\citet{JCC4:JCC4393} describe social network sites (SNSs) as: \begin{quote}``Web-based services that allow individuals to (1) construct a public or semi-public profile within a bounded system, (2) articulate a list of other users with whom they share a connection, and (3) view and traverse their list of connections and those made by others within the system.''\end{quote}

During the last decade, online social networks have become the outstanding source of multimedia information.

\subsection{Last.fm}
Last.fm is a social network system that accumulate a list of played audio tracks from registered users through \emph{scrobbling} to provide to any user a detail about listening preference and taste similarites between connected friends in the network. Last.fm also uses scrobbling to feed its music recommendation service to help to users to discover new artists.

Users' information such as recently played tracks, loved tracks, or top songs over a time period e.g. weeks, months, can be retrieved by using Last.fm API\footnote{http://www.last.fm/api} methods.

\section{Music services platforms}

\subsection{Echonest}
\subsection{7Digital}

Both Echo Nest and 7digital require to sign up to their API to get unique keys for OAuth authentication in order to retrieve desired information. As well, free account has limited number of calls, in the case of Echo Nest is limited to 20 request per minute and in the case of 7digital is limited to 4000 request per day.
%\subsection{APIs}
%The publicly available music related information can be collected from user profiles on social networks using Application Program Interface (API).

%\section{Data Fusion Techniques}
%Combination of multiple sources of information to obtain more relevant parameters is known as data fusion.
%In this study, a cooperative data fusion technique is considered to augment information provided from social network source to content-based system features. \citep{Castanedo2013}

\section{Recommender Systems}

Recommender systems are software or technical facilities to provide items suggestions or predict customer preferences. These systems play an important role in commercial applications to increase items sales and user satisfaction. In general, recommender systems can be categorised in the following groups: collaborative filtering and content-based methods. 

\subsection{Collaborative filtering}
In collaborative filtering (CF), recommendations are based on correlation between users' ratings or they can be predicted from historical user data. The strength of CF is that the recommendation process is independent from the item features. On the other hand, CF would not be suitable if the user-item matrix is sparse. \citep{Burke2002331}

\subsection{Content-based methods}
Content-based methods build user profiles by analysing the users' rated items. Each profile is then processed to be correlated with another item, which has not been rated, to compute the interest of the user on this object. \citep{Lops2011}

\section{Hybrid recommender methods}
Hybrid recommendation is based on the combination of techniques mentioned above, by using the advantages of one system to compensate the disadvantages of the other system. 

In this project, CF, that provides song ratings, is integrated with a content-based method, that compare spectral features of song to achieve hybridisation. 
%is based on a three-way aspect model \citep{Yoshii2008435}. Real item ratings are obtained through Last.fm API and spectral information are represented by convolutional deep belief networks (CDBN) features computed from items' spectrogram \citep{Lee20091096}.


\section{Music Information Retrieval}
Music Information Retrieval (MIR) is an extend of audio signal processing for understanding the usefulness and applications of music data by using time-frequency representations or low-level features. Applications of MIR include artist identification, genre classification and music recommendation.
\subsection{Musical genre classification}
Music classification is one of the principal components for clustering audio tracks based on similarities between features of pieces of music. Automatic musical genre classification approach proposed by \citet{Tzanetakis2002293}, which uses GTZAN genre dataset\footnote{http://marsyas.info/downloads/datasets.html}, has been widely used in the past decade. Nonetheless, the GTZAN dataset has inaccuracies \citep{Sturm20127}, it still provides an useful baseline to compare musical genre classification systems.

\subsection{Deep Learning}
One of the aims of learning algorithms is to identify high-level features that help us make sense of an observed data., e.g. genre, mood or release time in a music library. However, it could be difficult to compute these abstract features directly from audio waveforms. Deep learning can solve the difficulty of extracting high-level representations by expressing them in terms of simpler features, e.g. spectrograms. Deep learning allows the computer to build complex concepts out of simpler concepts. \citep{Bengio-et-al-2015-Book}

\citet{Sigtia20146959} examined and compared three implementations of deep neural networks to learn features for music genre classification, using Rectifier Linear Units (ReLUs), dropout regularisation and Hessian Free optimization.

\subsection{Convolutional Neural Networks}
Convolutional Neural Networks (CNNs) are type of neural networks that uses convolution operation instead of matrix multiplication for processing data that has grid-like topology \citep{Bengio-et-al-2015-Book} such as images collection.

\citet{NIPS2013_5004} used a convolutional network approach to predict latent factors from music audio in a content-based recommendation system.

\section{Estimation of Distribution Algorithms}
Estimation of distribution algorithms (EDAs) \citep{pelikan2015estimation} are optimisation techniques by constructing a probabilistic model from a sample of solutions, generating a new population and leading to an optimal solution \citep{Santana:Bielza:Larrañaga:Lozano:Echegoyen:Mendiburu:Armañanzas:Shakya:2009:JSSOBK:v35i07}.

\citet{Liang2014781} exploited an EDA to model user profiles by using weighted featured vectors of keywords from a set of items that the user had rated above a threshold.
\\
\\
In this chapter, previous work on recommender systems has been reviewed and novelty techniques for representing acoustical features and for modelling user profiles has been presented. The next step is to implement the algorithms to collect the dataset by crawling online social information, to extract the acoustical features of a collection of songs for representing them as vectors, to model the user profiles by an EDA, and therefore, to return predicted recommendations.