Please use this identifier to cite or link to this item:
Title: Gated graph neural networks for building session-based recommender systems
Keywords: session-based recommendation
user behavior analysis
gated graph neural networks
graph representation learning
Issue Date: 10-Jul-2023
Publisher: Ibn Khaldoun University
Abstract: In the modern world, consumers are faced with an overwhelming number of options and alternatives in various domains and applications. This phenomenon, known as choice overload, can cause confusion, dissatisfaction, and decision paralysis among consumers. Therefore, there is a growing demand for recommendation systems that can assist consumers in filtering out irrelevant options and guiding them to the most suitable ones based on their preferences and needs. Recommendation systems are software tools and techniques that provide personalized suggestions for items that users may like or need. They have become essential in various domains and applications, namey, e-commerce, streaming media, news platforms, and social media platforms. They can enhance user satisfaction, loyalty, and retention, as well as increase sales, conversion, and revenue for businesses. Session-based recommendation is a challenging task that aims to predict user actions based on anonymous sessions. Existing methods use sequential models to learn user and item representations for recommendation. However, these methods have limitations in capturing accurate user vectors in sessions and modeling complex item transitions. In this thesis, we propose a novel method called L-GGNN-ATT, which stands for Lossless Gated Graph Neural Networks with Attention as Readout. L-GGNN-ATT treats session sequences as graph-structured data and applies graph neural network – which is a deep learning method to infer on graphs – to learn item embeddings that can reflect complex item transitions. Moreover, it uses an attention network to represent each session as a combination of the global preference, i.e, the various item interactions, and the current interest of that session, which is simply the current item we are on. We conduct extensive experiments on two real datasets and demonstrate that our model significantly outperforms the state-of-the-art methods for session-based recommendation.
Appears in Collections:Master

Files in This Item:
File Description SizeFormat 
TH.M.INF.2023.32.pdf2,54 MBAdobe PDFView/Open

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.