Today, the way we listen to music has drastically changed compared to the past. With the advent of the internet and music streaming services, everyone has access to huge music libraries. This new scenario has created the necessity to develop Music Recommendation Systems. These are systems that guide users' music exploration analyzing their musical tastes and habits. Music exploration can be understood as a real discovery of contents, even not strictly belonging to one's musical tastes. Streaming services, such as Spotify, tend to propose tracks very connected to the user, preferring items with greater popularity, which is not always desired. Many users do not thoroughly understand the logic behind the recommendation generated by automatic systems, preferring instead to receive suggestions from their friends. People are generally willing to listen to songs far from their musical tastes if they are recommended by a friend, to create a social and shared experience. Starting from these considerations we develop a method for the creation of exploratory playlists that aim to create a musical path between two people, and therefore between their musical tastes. In order to realize it, a disentangled multidimensional music similarity space trained with a Conditional Similarity Network is used to represent compactly each song. Then, the users' music profiles are defined through their listening histories. Finally, the points in the space representing the songs of the final playlists are computed. Each user can choose the friend towards whom steer the exploration and eventually also a dimension of similarity between genre, mood, instrument, and era upon which the music embeddings space will be based. To evaluate the method, three experiments were performed. The first one concerns the training of the embedding extractor network. It was performed using offline and online triplet mining techniques. Another experiment was performed to test the playlist generation algorithm on the Last.fm publicly available dataset. Trends in tag presence were analyzed along playlists, to assess the smoothness of transitions between songs. In addition, we verified that selected songs were more or less equidistant in the embedding space. The last experiment involved evaluating user experience with real users. We used participants' music histories to generate personalized playlists for them. The evaluation is performed through two surveys asking about liking, exploration, and perception about the musical path toward the selected friend. The results of the first experiment regard the evaluation of the music space generated with the test set. It showed us that the best technique was to use a combination of two training strategies. The first uses offline triplets and the second uses an online triplet mining technique called batch semi-hard. The results of the second experiment confirmed that the trend of tag presence decrease with respect to the tags contained in the first song and increase with respect to the tags in the last song. Therefore, gradual similarity transition between the start and end points is realized. On the other hand, the evaluation of the last experiment confirms that the degree of exploration is very high. Despite this, the liking is good on average. The evaluations of the total experience are very positive and people found very interesting the possibility to choose a friend for the discovery of new music. Future works could examine other dimensions of similarity to include, and provide a user interface to support the method, with the aim to give a complete user experience. Moreover, it could be interesting to extend this method for group recommendation tasks.
Oggi il modo di ascoltare la musica è drasticamente cambiato rispetto al passato. Con l'avvento di Internet e dei servizi di streaming musicale ciascun utente ha la possibilità di accedere a enormi librerie musicali. Questo nuovo scenario ha reso necessario lo sviluppo di sistemi di raccomandazione musicale, sistemi che guidano l’esplorazione musicale degli utenti in base ai loro gusti musicali e alle loro abitudini. Con esplorazione musicale intendiamo una vera e propria scoperta di nuovi contenuti, che non appartengano necessariamente ai gusti musicali dell'utente. Servizi di streaming, come per esempio Spotify, tendono a proporre brani strettamente collegati all’utente e con maggiore popolarità, cosa non sempre desiderata. Molti utenti infatti non comprendono pienamente la logica dietro una determinata raccomandazione, preferendo invece ricevere consigli dai propri amici.Infatti generalmente sono disposti ad ascoltare canzoni anche lontane dai propri gusti musicali se vengono consigliate da un amico, creando un esperienza condivisa con quest’ultimo. Partendo da queste considerazioni, in questa tesi proponiamo un metodo per la creazione di playlist esplorative con l’obiettivo di tracciare un percorso musicale tra due persone, e quindi tra i loro gusti musicali. Per realizzarlo, ogni canzone viene rappresentata come un punto in uno spazio musicale multidimensionale di similarità, implementato tramite una Conditional Similarity Network. Il profilo musicale di ciascun utente viene definito analizzando la cronologia di ascolto. Infine vengono calcolati i punti nello spazio che rappresentano le canzoni finali che faranno parte delle playlist. Ogni utente ha la possibilità di scegliere l'amico verso cui dirigere l’esplorazione ed eventualmente anche una dimensione di similarità tra genre, mood, instrument ed era sulla quale sarà definito lo spazio musicale di partenza. Per valutare il metodo abbiamo eseguito tre diversi esperimenti. Nel primo esperimento analizziamo il training della rete e valutiamo la qualità dello spazio creato. Il training è stato eseguito utilizzando tecniche di estrazione di terzine offline e online. Un secondo esperimento è stato effettuato per valutare l'algoritmo di generazione delle playlist usando un dataset pubblico, Last.fm dataset 1k. Per valutare la fluidità delle transizioni tra le canzoni sono stati analizzati gli andamenti sulla presenza dei tag lungo le playlist. Nell'ultimo esperimento valutiamo l'esperienza utente complessiva, usando le cronologie musicali dei partecipanti per generare playlist personalizzate. Le valutazioni sono state ottenute tramite la compilazione di due sondaggi riguardanti il gradimento, l'esplorazione e la percezione del percorso musicale verso l'amico selezionato. I risultati del primo esperimento ci hanno mostrato che la strategia migliore è utilizzare una combinazione di due training, cioè partendo con la creazione offline di triplette e proseguendo con una tecnica di calcolo di triplette online, chiamata batch semi-hard. I risultati del secondo esperimento hanno confermato che l'andamento della presenza dei tag è decrescente rispetto ai tag contenuti nella prima canzone e crescente rispetto a quelli dell'ultima canzone. Questo significa che il sistema di generazione della playlist è in grado di realizzare un percorso graduale tra il punto di partenza e quello di arrivo. La valutazione dell'ultimo esperimento conferma invece che il grado di esplorazione è molto alto e, nonostante questo, il gradimento in media è buono. Le valutazioni dell'esperienza complessiva sono molto positive: in particolar modo le persone hanno trovato molto interessante la possibilità di scegliere un amico per guidare l'esperienza di esplorazione musicale. Lavori futuri potrebbero riguardare l'utilizzo di altre dimensioni di similarità musicale e prevedere un interfaccia utente a supporto del metodo, creando quindi un'esperienza utente completa. Inoltre potrebbe essere interessante estendere questo metodo per creare playlist a partire dai gusti di un gruppo di utenti.
Playlist generation for music exploration shared between users using conditional similarity networks
Pulvirenti, Carlo
2020/2021
Abstract
Today, the way we listen to music has drastically changed compared to the past. With the advent of the internet and music streaming services, everyone has access to huge music libraries. This new scenario has created the necessity to develop Music Recommendation Systems. These are systems that guide users' music exploration analyzing their musical tastes and habits. Music exploration can be understood as a real discovery of contents, even not strictly belonging to one's musical tastes. Streaming services, such as Spotify, tend to propose tracks very connected to the user, preferring items with greater popularity, which is not always desired. Many users do not thoroughly understand the logic behind the recommendation generated by automatic systems, preferring instead to receive suggestions from their friends. People are generally willing to listen to songs far from their musical tastes if they are recommended by a friend, to create a social and shared experience. Starting from these considerations we develop a method for the creation of exploratory playlists that aim to create a musical path between two people, and therefore between their musical tastes. In order to realize it, a disentangled multidimensional music similarity space trained with a Conditional Similarity Network is used to represent compactly each song. Then, the users' music profiles are defined through their listening histories. Finally, the points in the space representing the songs of the final playlists are computed. Each user can choose the friend towards whom steer the exploration and eventually also a dimension of similarity between genre, mood, instrument, and era upon which the music embeddings space will be based. To evaluate the method, three experiments were performed. The first one concerns the training of the embedding extractor network. It was performed using offline and online triplet mining techniques. Another experiment was performed to test the playlist generation algorithm on the Last.fm publicly available dataset. Trends in tag presence were analyzed along playlists, to assess the smoothness of transitions between songs. In addition, we verified that selected songs were more or less equidistant in the embedding space. The last experiment involved evaluating user experience with real users. We used participants' music histories to generate personalized playlists for them. The evaluation is performed through two surveys asking about liking, exploration, and perception about the musical path toward the selected friend. The results of the first experiment regard the evaluation of the music space generated with the test set. It showed us that the best technique was to use a combination of two training strategies. The first uses offline triplets and the second uses an online triplet mining technique called batch semi-hard. The results of the second experiment confirmed that the trend of tag presence decrease with respect to the tags contained in the first song and increase with respect to the tags in the last song. Therefore, gradual similarity transition between the start and end points is realized. On the other hand, the evaluation of the last experiment confirms that the degree of exploration is very high. Despite this, the liking is good on average. The evaluations of the total experience are very positive and people found very interesting the possibility to choose a friend for the discovery of new music. Future works could examine other dimensions of similarity to include, and provide a user interface to support the method, with the aim to give a complete user experience. Moreover, it could be interesting to extend this method for group recommendation tasks.File | Dimensione | Formato | |
---|---|---|---|
Thesis-Pulvirenti.pdf
Open Access dal 03/04/2023
Descrizione: Tesi di laurea magistrale
Dimensione
36.35 MB
Formato
Adobe PDF
|
36.35 MB | Adobe PDF | Visualizza/Apri |
Executive_Summary-Pulvirenti.pdf
Open Access dal 03/04/2023
Descrizione: Executive Summary
Dimensione
4.92 MB
Formato
Adobe PDF
|
4.92 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/186399