In personalized e-commerce recommendation systems, it's very hard to guess what the next click will be during an active user session because there are very few interactions (more than 90%), user intent changes frequently, and item-to-item relationships are complicated. In very fast-paced retail settings, current session-aware models don't work very well because they only look at short-term intent or long-term preference modeling. The Contrastive Session Pattern Mixer (CSPM) model developed in this work addresses personalized e-commerce recommendations and session-aware next-click predictions. The CSPM architecture combines a session pattern mixer with contrastive representation learning to mimic both the dynamics inside a session and the similarities between sessions. Multi-granular pattern mixing layers improve intent-awareness by roughly 20% compared to standard session-based architectures. This is because they capture both local item transitions and global behavioral structures. A contrastive learning target aligns semantically similar sessions while separating dissimilar ones. This makes representation more resilient and allows for better generalization with roughly 25% fewer interaction samples.According to a lot of tests, CSPM consistently outperforms the best session-based recommendation models on benchmark e-commerce clickstream datasets. CSPM raises the Hit Rate by 12–15%, the NDCG by 13–17%, and the Mean Reciprocal Rank (MRR) by 11–14%. Another good thing about contrastive session alignment is that it makes predictions more stable by more than 18%, especially in sparse or noisy interaction sessions. CSPM doesn't need clear user profiles to quickly figure out what users want to do, which makes it a good choice for real-world e-commerce systems that need to be able to grow. The suggested approach greatly increases the accuracy of session-aware suggestions, making it very useful for big, dynamic online retail systems.