Improvements in Recommendation Systems Using Graph Neural Networks
Akhilesh Kumar Srivastava, Ishanki Verma, Puneet Garg
Abstract
The concept of graph neural networks has shown improvements in many applications due to its ability to handle data involving complex relationships. As the population demonstrates a consistent upward trend, concomitant with this demographic expansion is the accrual of an augmented dataset. This project involves the use of graph neural networks in order to handle expanding data. Large dataset requires effective handling and so the concept of multiprocessing is used in order to handle data with maximum utilisation of CPU. Random walk is used for generating node embeddings for the graph. Node embeddings are then utilised for link prediction which is used for assessing the likelihood of relationships between the nodes. There is also an option of leveraging metapath schemes in order to guide the random walk. The key concepts also include variation of N ode2vec algorithm, use of neighbour sampling, use of negative sampling during training and noise contrastive estimation(NCE) as a loss function.