Recommender systems became part of everyday life. They are present when choosing what movie to watch on Netflix, what book to buy on Amazon, and even to find friends on Facebook. Progressively, people are turning to these systems to help them interact more effectively with overwhelming amounts of content, and find the information that is most valuable to them. Deep learning-based recommendation systems has increased in the past recent years to overcome the limitations of traditional models, specially when dealing with the enormous volume, complexity and dynamics of data. One successful application is to consider the collaborative filtering from an autoencoder perspective in order to achieve better results. The collaborative denoising autoencoder (CDAE) is a flexible top-n recommender that uses a corrupted version of the user-item interactions to make recommendations. The type of process used to perform the corruption is an important aspect that substantially affects the final representation learned from the input, and therefore has an impact in the performance. There are two core purposes for this work. To begin with, it is aimed to replicate the results of the experiments for the collaborative denoising autoencoder (CDAE) model proposed by Wu Y., DuBois C., Zheng A. X. and Ester M. [13], so as to have a clear starting point for further analysis and be able to understand the components of this recommender. Moreover, it proposes novel ways to introduce noise to the input data supplied to the CDAE aiming to overcome the issues inherent in state-of-the-art choices for corruption and improve performance. Usually, in such traditional approaches the noise is stochastically applied to the input, either by some additive mechanism or randomly masking some of the input values, and so, it is hard to have control over it.
I sistemi di suggerimento sono diventati parte della vita quotidiana. Sono presenti quando si sceglie quale film guardare su Netflix, quale libro acquistare su Amazon e persino per trovare amici su Facebook. Progressivamente, le persone si rivolgono a questi sistemi per aiutarli a interagire in modo piu` efficace con una quantit`a enorme di contenuti e trovare le informazioni piu` preziose per loro. I sistemi basati sul deep learning sono aumentati negli ultimi anni per superare i limiti dei modelli tradizionali, specialmente quando si ha a che fare con l’enorme volume, complessit`a e dinamica dei dati. Un’applicazione di successo consiste nel considerare il filtro collaborativo da una prospettiva di autoencoder al fine di ottenere risultati migliori. Il collaborative denoising autoencoder (CDAE) `e un top-n recommender flessibile che utilizza una versione danneggiata delle interazioni tra l’utente e gli oggetti per formulare raccomandazioni. Il tipo di processo utilizzato per eseguire la corruzione `e un aspetto importante che ha impatto importante sulla rappresentazione finale appresa dall’input e quindi ha rilevanza sulla performance. Ci sono due scopi principali per questo lavoro. Per cominciare, ha la finalit`a di replicare i risultati degli esperimenti per il collaborative denoising autoencoder (CDAE) proposto da Wu Y., DuBois C., Zheng AX e Ester M. [13], in modo da avere un chiaro punto di partenza per ulteriori analisi e poter comprendere i componenti di questo suggeritore. Inoltre, propone nuovi modi per introdurre il rumore nei dati di input forniti al CDAE con l’obiettivo di superare i problemi inerenti alle scelte all’avanguardia per la corruzione. Di solito, in tali approcci tradizionali il rumore viene applicato in modo stocastico all’input, o mediante un meccanismo additivo o mascherando casualmente alcuni dei valori di input, e quindi `e difficile avere il controllo al riguardo.
Non-traditional approaches to input corruption in collaborative denoising autoencoders
SCHIATTI, LAURA
2019/2020
Abstract
Recommender systems became part of everyday life. They are present when choosing what movie to watch on Netflix, what book to buy on Amazon, and even to find friends on Facebook. Progressively, people are turning to these systems to help them interact more effectively with overwhelming amounts of content, and find the information that is most valuable to them. Deep learning-based recommendation systems has increased in the past recent years to overcome the limitations of traditional models, specially when dealing with the enormous volume, complexity and dynamics of data. One successful application is to consider the collaborative filtering from an autoencoder perspective in order to achieve better results. The collaborative denoising autoencoder (CDAE) is a flexible top-n recommender that uses a corrupted version of the user-item interactions to make recommendations. The type of process used to perform the corruption is an important aspect that substantially affects the final representation learned from the input, and therefore has an impact in the performance. There are two core purposes for this work. To begin with, it is aimed to replicate the results of the experiments for the collaborative denoising autoencoder (CDAE) model proposed by Wu Y., DuBois C., Zheng A. X. and Ester M. [13], so as to have a clear starting point for further analysis and be able to understand the components of this recommender. Moreover, it proposes novel ways to introduce noise to the input data supplied to the CDAE aiming to overcome the issues inherent in state-of-the-art choices for corruption and improve performance. Usually, in such traditional approaches the noise is stochastically applied to the input, either by some additive mechanism or randomly masking some of the input values, and so, it is hard to have control over it.| File | Dimensione | Formato | |
|---|---|---|---|
|
Non_traditional_approaches_to_input_corruption_in_Collaborative_Denoising_Autoencoders.pdf
accessibile in internet per tutti
Dimensione
980.92 kB
Formato
Adobe PDF
|
980.92 kB | 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/10589/170595