In this history time frame every single individual produce an immense amount of data encompassing their visual experiences, dietary choices, viewing preferences, and various other personal interests. This wealth of information presents a valuable opportunity for exploitation by large organizations, which invest considerable time and resources into specialized information filtering systems known as recommender systems. Thanks to these particular programs companies are able to offer a really useful service that could predict every kind of digital content or item according to the tastes of the final user. There are mainly 2 categories of these models: collaborative based and content based. The first one focuses on the preferences and behaviors of similar users in the past to provide recommendations while the other relies on the characteristics of the items in order to find similar ones. In the majority of the cases is really hard to provide efficient and fast systems due to the lack of coherent and valid data, in addition to this the interactions that come from the users are influenced by many biases. In this thesis we aim to develop novel collaborative deep learning recommender algorithms combined with specific architectures in order to advance the state of the art and improve its quality by benchmarking against other similar systems focused on the same goal or using the same approaches.
In questo periodo storico praticamente ogni persona produce una quantità immensa di dati che possono rappresentare ciò che vede, ciò che mangia, ciò che gli piace guardare e molte altre informazioni che caratterizzano i suoi gusti. Per sfruttare tutte queste preziose statistiche le grandi organizzazioni investono molto tempo e denaro in particolari architetture di filtraggio delle informazioni chiamate sistemi di raccomandazione. Grazie a questi particolari programmi, le aziende sono in grado di offrire un servizio veramente utile che può prevedere ogni tipo di contenuto digitale o articolo preferito a seconda dei gusti dell’utente finale. Ci sono principalmente due categorie di modelli: basati sul collaborative filtering e sul content filtering. Il primo si concentra sulle preferenze e i comportamenti di utenti simili avvenuti in precedenza per fornire raccomandazioni, mentre l’altro si basa sulle caratter istiche principali degli articoli per consigliarne di nuovi. Nella maggior parte dei casi è molto difficile fornire sistemi efficienti e veloci a causa della mancanza di dati coerenti e validi; inoltre, le interazioni che provengono dagli utenti sono influenzate da molti bias. In questa tesi, ci proponiamo di sviluppare nuovi algoritmi di raccomandazione collabo rativa basati su algoritmi di deep learning combinati con architetture specifiche al fine di progredire lo stato dell’arte e migliorarne la qualità confrontandoli con altri sistemi simili focalizzati sullo stesso obiettivo o che utilizzano gli stessi approcci.
Comparative Analysis Of The State Of The Art For Collaborative Filtering: Enhancing LightGCN With Attention Based Graph Neural Networks And Positional Embeddings
BRUNELLO, SIMONE
2023/2024
Abstract
In this history time frame every single individual produce an immense amount of data encompassing their visual experiences, dietary choices, viewing preferences, and various other personal interests. This wealth of information presents a valuable opportunity for exploitation by large organizations, which invest considerable time and resources into specialized information filtering systems known as recommender systems. Thanks to these particular programs companies are able to offer a really useful service that could predict every kind of digital content or item according to the tastes of the final user. There are mainly 2 categories of these models: collaborative based and content based. The first one focuses on the preferences and behaviors of similar users in the past to provide recommendations while the other relies on the characteristics of the items in order to find similar ones. In the majority of the cases is really hard to provide efficient and fast systems due to the lack of coherent and valid data, in addition to this the interactions that come from the users are influenced by many biases. In this thesis we aim to develop novel collaborative deep learning recommender algorithms combined with specific architectures in order to advance the state of the art and improve its quality by benchmarking against other similar systems focused on the same goal or using the same approaches.| File | Dimensione | Formato | |
|---|---|---|---|
|
2024_07_Brunello_Executive Summary.pdf
accessibile in internet per tutti
Descrizione: Executive Summary
Dimensione
691.12 kB
Formato
Adobe PDF
|
691.12 kB | Adobe PDF | Visualizza/Apri |
|
2024_07_Brunello_Tesi.pdf
accessibile in internet per tutti
Descrizione: Tesi
Dimensione
2.2 MB
Formato
Adobe PDF
|
2.2 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/10589/223396