Lightweight Transformer-Based Event Embeddings with Log-Frequency Session Weighting for Purchase Prediction in E-Commerce Clickstream Data
Workshops of 26th International Conference on Computational Science and its Applications, ICCSA 2026, Braga, Portekiz, 30 Haziran - 03 Temmuz 2026, cilt.16759 LNCS, ss.242-259, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Cilt numarası: 16759 LNCS
- Doi Numarası: 10.1007/978-3-032-30518-3_16
- Basıldığı Şehir: Braga
- Basıldığı Ülke: Portekiz
- Sayfa Sayıları: ss.242-259
- Anahtar Kelimeler: Clickstream Data Analysis, Event Embeddings, Frequency Weighted Aggregation, Lightweight Transformers, Purchase Prediction, Session Representation
- Yıldız Teknik Üniversitesi Adresli: Evet
Özet
E-commerce platforms generate large volumes of clickstream data in the form of short user sessions. Predicting whether the next event in a session will be a purchase is a useful task for recommendation, marketing analytics, and real-time decision support. Many existing methods address this problem with complex sequential deep learning models, but such models often require substantial computational resources and large training datasets. This limits their practicality in resource-constrained settings. In this work, we study a lightweight alternative based on event embeddings and frequency-aware session representation. Each clickstream event token, such as a page view or user action, is encoded as a dense vector using a lightweight transformer-based embedding model. To reflect repeated behavior within a session, we apply a logarithmic frequency-weighting scheme so that recurring events contribute more strongly to the final session representation without dominating it. The weighted event embeddings are then aggregated into a fixed-length session vector and used as input to supervised classifiers for next-event purchase prediction. The approach is evaluated on both a public clickstream dataset and a large real-world e-commerce dataset. Experimental results show that frequency-weighted session representations perform better than unweighted representations across different embedding models and classifiers. The gains are more evident in highly imbalanced settings where purchase events are rare. These findings indicate that lightweight transformer-based embeddings combined with log-frequency weighting offer a practical approach for next-event purchase prediction in large-scale clickstream environments.