Balancing Recency and Order in Session Representations: A Temporal-Sequential Embedding Approach for Clickstream Analytics
2025 IEEE International Conference on Big Data, BigData 2025, Macau, Çin, 8 - 11 Aralık 2025, ss.5362-5372, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/bigdata66926.2025.11401493
- Basıldığı Şehir: Macau
- Basıldığı Ülke: Çin
- Sayfa Sayıları: ss.5362-5372
- Anahtar Kelimeler: Clickstream analytics, hybrid embeddings, purchase propensity prediction, temporal and recency-based modeling, user behavior analysis
- Yıldız Teknik Üniversitesi Adresli: Evet
Özet
This study developed and evaluated an embeddingbased framework for real-time prediction of purchase propensity in large-scale e-commerce clickstream data. The proposed system generated compact and discriminative session representations using three strategies: a time-weighted embedding emphasizing recent actions, a recency-based averaging approach focusing on short-term behavior, and a sequence-based neural representation capturing event order. The motivation was to explore whether recency-oriented or order-sensitive embeddings more effectively represent evolving user intent in session-based modeling. Experimental analyses using multiple classifiers under stratified validation demonstrated that the time-weighted approach achieved the strongest predictive performance, while hybrid combinations integrating temporal and sequential features provided additional gains. All methods operated efficiently, producing embeddings within milliseconds and enabling real-time applicability. The findings highlight that temporally weighted embeddings offer the best balance between accuracy, scalability, and early intent detection, contributing to more adaptive and interpretable clickstreambased recommendation systems.