Hibridno priporočanje glasbe z nevronskimi mrežami na grafih

diplomsko delo

Matej Bevec (Author), Matevž Pesek (Mentor), Marko Tkalčič (Co-mentor)

Abstract

Odkar so poslušalske platforme, kot je Spotify, postale prevladujoče sredstvo glasbene potrošnje, se le te zanašajo na priporočilne sisteme, da uporabnike pomagajo usmerjati po svojih vse obsežnejših glasbenih zbirkah. V preteklosti so bile pri tej nalogi najuspešnejše metode sodelovalnega filtriranja (angl. collaborative filtering, CF), ki izkoriščajo pretekle interakcije med uporabniki in artikli. Vseeno pa te metode niso brez pomanjkljivosti. Pogosto so npr. znatno manj uspešne med slabše poznanimi artikli z le malo podatki o interakcijah. Nasprotno modeli, ki temeljijo na vsebini, ta problem popolnoma zaobidejo, saj tvorijo priporočila izključno na osnovi vsebine artiklov (npr. zvočnega signala skladb), vendar pa take metode še niso primerljivo zmogljive. Strojno učenje z grafi --- rastoče področje, ki prekaša pretekle pristope v mnogih domenah --- morda lahko naslovi tudi tukajšno nalogo. Zlasti najnovejše metode, nevronske mreže na grafih (angl. graph neural networks, GNN), obljubljajo, da lahko učinkovito izkoristijo tako uporabniške informacije v obliki grafa kot tudi vsebino in tako generirajo hibridna priporočila. V pričujočem delu sodoben algoritem GNN, PinSage, v vlogi hibridnega priporočilnika apliciramo na javne podatke platforme Spotify. Algoritem naučimo na novem grafu vključenosti skladb v seznamih predvajanja in ga evalviramo v primerjavi z metodami na osnovi vsebine, matričnimi metodami CF in metodami CF na osnovi grafov, pri čemer obravnavamo tudi kriterije onkraj natančnosti (angl. beyond accuracy objectives). Naši eksperimenti kažejo sledeče. Implementirani algoritem je med najuspešnejšimi in izstopa kot najbolj robusten. Metode CF na osnovi grafov so bistveno boljše od matričnih metod CF, kar nakazuje, da je uporabniške informacije bolje predstaviti kot graf. V okviru evalvacijske naloge metode CF ne kažejo padca učinkovitosti v dolgem repu, kjer hibridni pristop posledično ne ponuja prednosti. Na podlagi rezultatov zaključimo, da je strojno učenje z grafi obetavna nova paradigma priporočanja glasbe, ki pa zahteva nekatere nadaljnje raziskave.

Keywords

priporočilni sistemi za glasbo;nevronske mreže na grafih;hibridni priporočilniki;vložitve;računalništvo in informatika;univerzitetni študij;diplomske naloge;

Data

Language:	Slovenian
Year of publishing:	2022
Typology:	2.11 - Undergraduate Thesis
Organization:	UL FRI - Faculty of Computer and Information Science
Publisher:	[M. Bevec]
UDC:	004.032.26:78(043.2)
COBISS:	119815683
Views:	44
Downloads:	8
Average score:	0 (0 votes)
Metadata:

Other data

Secondary language:	English
Secondary title:	Hybrid music recommendation with graph neural networks
Secondary abstract:	Since streaming has emerged as the predominant means of music consumption, platforms such as Spotify rely on recommender systems to help users navigate within their increasingly large song catalogues. Collaborative filtering (CF) methods, which rely on past user-item interactions, have historically been most successful. They do, however, have various limitations, like performing poorly among lesser-known items with little or no interaction data. Conversely, content-based models attempt to circumvent the data-sparsity issue by generating recommendations based on item content (song audio) alone, but have seen limited success. A growing paradigm, which is seeing significant success in other fields and may help address the task at hand, is graph-based ML. Graph neural networks (GNN) in particular, promise to learn from both the complex relationships within a user-interaction graph, as well as from content to generate hybrid recommendations. In our work, we introduce graph-based ML to the field of music recommendation by applying a state-of-the-art GNN, PinSage, to public Spotify data. We train the implemented algorithm on a newly collected playlist-song membership graph and evaluate it against traditional CF, graph-based CF and content-based methods on a related-song prediction task, venturing beyond-accuracy in our evaluation. Our experiments show the following. Our approach is among top performers and stands out as the most well rounded. Graph-based CF methods significantly outperform matrix-based CF approaches, suggesting that user interaction data may be better represented as a graph. In the scope of our evaluation task, CF methods do not exhibit a performance drop in the long tail, where the hybrid approach does not offer an advantage. Based on our results we conclude that, although requiring further research, graph-based ML is a promising future direction for music recommendation.
Secondary keywords:	music recommender systems;graph neural networks;hybrid recommenders;embeddings;computer science;computer and information science;diploma;Glasba;Nevronske mreže (računalništvo);Računalništvo;Univerzitetna in visokošolska dela;
Type (COBISS):	Bachelor thesis/paper
Study programme:	1000468
Embargo end date (OpenAIRE):	1970-01-01
Thesis comment:	Univ. v Ljubljani, Fak. za računalništvo in informatiko
Pages:	67 str.
ID:	16306617