Purpose: Session-based recommendation using graph neural networks (GNN) is a popular approach to model users' behaviors and attributes of items from the perspective of user-item interaction sequence. However, current researches seldom incorporate the unique attributes of items to delve into a comprehensive analysis of user behaviors. In addition, GNN faces three problems when encountering complex modeling scenarios: long-range dependencies, order information loss, and data sparsity, which are essential to modeling long-tail items. Methods: We study the interactions between users and items from a new perspective. A novel Contrastive Learning based Tail Adjusted Repeat Aware Graph Neural Network (CLTAR-GNN) is proposed to tackle the problems. A Tail Adjusted Repeat (TAR) mechanism captures users' repeat-explore behaviors in both short-head and long-tail session items based on graph neural networks. Through the TAR, we are able to further understand the underlying graph-based mechanisms that influence user-item interactions. A Self-Attention (SA) network with position embedding is incorporated to overcome the sequence information loss issues, which may be caused by the complex user behaviors and item characteristics modeling. Finally, a mutli-task learning framework is employed to combine TAR, SA and a contrastive learning model into a unified framework to enhance model performance by collaboratively training graph and sequence-based embeddings. Results: Experimental results show that CLTAR-GNN outperforms the state-of-the-art session-based recommendation methods significantly. The average improvement compared with all baselines are 17.5% (HR@20) and 22.5% (MRR@20) on both experimental datasets.