Abstract
Session-based recommendation is a challenging problem due to the inherent uncertainty of user behavior and the limited historical click information. Latent factors and the complex dependencies within the user’s current session have an important impact on the user's main intention, but the existing methods do not explicitly consider this point. In this paper, we propose a novel model, Interest Shift and Latent Factors Combination Model (ISLF), which can capture the user's main intention by taking into account the user’s interest shift (i.e. long-term and short-term interest) and latent factors simultaneously. In addition, we experimentally give an explicit explanation of this combination in our ISLF. Our experimental results on three benchmark datasets show that our model achieves state-of-the-art performance on all test datasets.
Introduction
Most recommendation engines rely on extensive long-term user histories. However, Session-based Recommendation (SRS) deals with anonymous users where only the current "click-stream" is known. This presents two primary challenges:
- Uncertainty and Noise: Users often click items by accident or out of idle curiosity, which does not reflect their true main intention.
- Interest Shift: A user’s goal can evolve during a session. They might start by looking at a broad category (long-term session interest) but end up fixated on a specific item (short-term interest).
Existing models often struggle to capture the underlying "why" behind a click—the hidden or latent factors—and fail to balance the user's overall session goal with their most recent actions.
Methodology: The ISLF Model
The Interest Shift and Latent Factors Combination Model (ISLF) is a hybrid architecture that treats recommendation as a combination of deep learning and probabilistic modeling.
The model consists of several key components designed to address the challenges of SRS:
- The Recurrent VAE: Instead of a standard RNN, ISLF uses a Variational Autoencoder (VAE) at each timestamp. The VAE infers "latent factor variables" that represent complex, hidden relationships between items, helping the model filter out noisy clicks.
- Long-term Interest Generator: This component utilizes a novel attention mechanism that considers both the order and frequency of clicks. If a user clicks an item multiple times, the model recognizes this repetition as a sign of high importance.
- Short-term Interest: This is represented by the embedding of the very last item clicked, capturing the user's immediate focus.
- The Hybrid Predictor: Finally, the model merges the latent factors, the long-term session representation, and the short-term interest into a single "overall interest" vector to predict the next item.
Experimental Results
The model was evaluated on two gold-standard datasets for session-based recommendation: YOOCHOOSE (using 1/64 and 1/4 fractions) and DIGINETICA.
Key Performance Metrics
ISLF outperformed several heavy-hitting baselines, including GRU-Rec, NARM, and STAMP:
- Recall@20: On the YOO-1/64 dataset, ISLF achieved 69.32%, a notable improvement over STAMP (68.74%) and NARM (68.32%).
- MRR@20 (Mean Reciprocal Rank): ISLF saw a significant jump here, scoring 33.58 on YOO-1/64, compared to STAMP’s 29.67. This suggests that ISLF is more effective at placing the correct item closer to the top of the recommendation list.
- Consistency: The model maintained superior performance across both short and long sessions, whereas many baselines tend to degrade as session length increases.
Discussion and Insights
- Latent Factors as a Core Driver: The inclusion of the VAE allows the model to capture hidden causal relationships. The authors found that latent factors alone provide a significant performance boost over standard RNNs.
- The Importance of Frequency: ISLF is one of the few models to explicitly use click frequency as a weight in its attention mechanism, proving that repetition is a strong signal of intent.
- Balancing Long and Short-term Interests: By combining a global view of the session with the local view of the last click, the model handles both steady browsing and sudden pivots in user interest.
Conclusion
ISLF demonstrates that explicitly modeling latent factors and interest shifts significantly improves session-based recommendations. While the model currently focuses on click sequences, future research aims to integrate auxiliary data like item categories and prices to further refine latent factor discovery and address the computational complexity of the VAE-RNN integration.
Summary updated