Fairness in Variational Autoencoders Recommenders
Imtiaz, Waleed (2020)
2020
Laskennallisen suurten tietoaineistojen analysoinnin maisterikoulutus, FM (engl) - Master's Degree Programme in Computational Big Data Analytics
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2020-04-30
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202004294462
https://urn.fi/URN:NBN:fi:tuni-202004294462
Tiivistelmä
Recommender systems have become increasingly relevant since big giants like Amazon, Netflix, and many other web services have started to enhance user’s experience based on recommendations. However, implementing state of the art system resulting in fair recommendations is the need of the hour.
Variational autoencoders (VAE) have proven to be very efficient to model user preferences in Collaborative Filtering space. Recently, the recurrent version of VAE that is based on the recurrent neural network has been proposed, known as Sequential Variational Autoencoder (SVAE), which includes temporal dependencies, unlike in simple VAEs. The predicted scores for a given set of items generated by SVAE is almost the same for top-ranked items in the recommendation list, which increase unfair treatment in the long term.
This thesis presents an overview of advanced recommender systems combined with historical and the state of the art methods used to carry out recommendation tasks efficiently. Furthermore, we also described unfairness as a problem for better recommendations and discuss ways to enhance fairness by incorporating random noises during the evaluation phase of SVAE. This randomness will provide long term fairness and make the given model fairness-aware, despite little loss in ranking quality Normalized Discounted Cumulative Gain (NDCG), which is tolerable. In the end, we also compared the value of fairness with respect to the loss in NDCG when we introduce different noise distributions. It turns out that its effect on ranking quality is much less as compared to the fairness we have achieved.
Variational autoencoders (VAE) have proven to be very efficient to model user preferences in Collaborative Filtering space. Recently, the recurrent version of VAE that is based on the recurrent neural network has been proposed, known as Sequential Variational Autoencoder (SVAE), which includes temporal dependencies, unlike in simple VAEs. The predicted scores for a given set of items generated by SVAE is almost the same for top-ranked items in the recommendation list, which increase unfair treatment in the long term.
This thesis presents an overview of advanced recommender systems combined with historical and the state of the art methods used to carry out recommendation tasks efficiently. Furthermore, we also described unfairness as a problem for better recommendations and discuss ways to enhance fairness by incorporating random noises during the evaluation phase of SVAE. This randomness will provide long term fairness and make the given model fairness-aware, despite little loss in ranking quality Normalized Discounted Cumulative Gain (NDCG), which is tolerable. In the end, we also compared the value of fairness with respect to the loss in NDCG when we introduce different noise distributions. It turns out that its effect on ranking quality is much less as compared to the fairness we have achieved.