Comprehending user behavior on e-commerce platforms is essential for augmenting customer interaction and refining recommendation algorithms. Clickstream data provides a significant resource for examining user navigation patterns; nevertheless, accurately describing and categorizing user sessions poses a challenge. This paper shows how to improve the accuracy of session-based classification using an embedding-based method combined with an aging-based weighting mechanism. The considered embedding methods, Word2Vec, Node2Vec, and LSTM Autoencoder, can turn session-based clickstream data into numbers. Furthermore, a dynamic weighting technique is introduced to emphasize recent interactions to improve classification performance. Our empirical assessment on an authentic e-commerce dataset reveals that the LSTM Autoencoder surpasses conventional embedding methods in capturing sequential dependencies. In addition, the age-based weighting technique markedly improves the classification accuracy, especially when used with deep learning models. A comparison of different classification algorithms, such as Random Forest, Logistic Regression, Gaussian Naive Bayes, and LSTM, shows that LSTM models are the best at finding correlations between events over time. The results also show the importance of temporal weighting in session-based clickstream analysis and provide a solid foundation for further research in behavioral analytics and personalized recommendation systems. This paper introduces an efficient method for clickstream-based user modeling that facilitates better user engagement in e-commerce systems.