Similarity-Based Supervised User Session Segmentation Method for Behavior Logs

TL;DR

This paper tackles the problem of dynamic within-session user interests by proposing a supervised session segmentation method that leverages similarity features derived from item embeddings and item attributes. It constructs features from a local window around each candidate segmentation point using behavior-based Item2Vec embeddings and text embeddings for titles and brands, plus a price-based similarity, and trains classifiers (with LightGBM delivering the best results). The approach outperforms a cosine-similarity baseline in F1 and PR-AUC across window sizes, with SHAP analyses showing title similarity as the most influential feature. The method serves as a robust preprocessing step for downstream session-based recommendations and behavioral analyses in e-commerce, while acknowledging limitations such as the need for manual annotations and non-personalized segmentation. Future work includes semi-supervised learning to reduce annotation costs and evaluating the segmentation impact on downstream predictive tasks.

Abstract

In information recommendation, a session refers to a sequence of user actions within a specific time frame. Session-based recommender systems aim to capture short-term preferences and generate relevant recommendations. However, user interests may shift even within a session, making appropriate segmentation essential for modeling dynamic behaviors. In this study, we propose a supervised session segmentation method based on similarity features derived from action embeddings and attributes. We compute the similarity scores between items within a fixed-size window around each candidate segmentation point, using item co-occurrence embeddings, text embeddings of titles and brands, and price information as sources for these similarity features. These features are used to train supervised classification models to predict the session boundaries. We construct a manually annotated dataset from real browsing histories and evaluate the segmentation performance using F1-score, PR-AUC, and ROC-AUC. The LightGBM model achieves the best performance, with an F1-score of 0.806 and a PR-AUC of 0.831. These results demonstrate the effectiveness of the proposed method for session segmentation and its potential to capture dynamic user behaviors.

Figures (4)

• Figure 1: Example of feature extraction for window size $w = 2$
• Figure 2: Distribution of session lengths
• Figure 3: Experiment flow
• Figure 4: Feature importance comparison across models