Session-based recommendation systems that work in a session’s context hold a lot of relevance in fields such as streaming services and e-commerce, where user interactions are recorded in sessions without taking into account their past behavior. However, there are existing models which struggle in precisely modeling the complex user and items interaction, temporal dependencies, and maintaining stable training in dynamic environments. This paper tends to solve these problems with the help of an architecture which fuses graph neural networks and advanced attention mechanisms. This model implements multi-head target attention aiming at diverse feature extraction, residual normalization for stable training, and positional masking to emphasize temporal relationship. We performed simulation on the proposed model with the help of two popular datasets named Diginetica and Yoochoose. We used MRR@20 and Recall@20 as evaluation metrics and have been able to outperform state-of-the-art session-based recommendation methods.