Recommender systems are an application of social network analysis that model user–item interactions. Session-based recommender systems make use of short-term user activities to generate personalized recommendations, but their performance can be affected by anomalous points due to noise, error, or unusual user behavior. This work extends a Graph Neural Network (GNN)-inspired framework by adding four anomaly detection methods: Isolation Forest, Local Outlier Factor (LOF), One-Class Support Vector Machine (SVM), and Density-Based Spatial Clustering of Applications with Noise (DBSCAN). These methods are applied at the session representation level and evaluated on three benchmark datasets (Yoochoose 1/64, Gowalla, and Diginetica). Experimental results indicate that session-level anomaly detection, particularly with moderately sparse datasets, significantly improves hit rate (HR) and mean reciprocal rank (MRR) and offers a reasonable improvement over baseline research.