Abstract
Sequential recommendation plays a vital role in modern recommendation systems by capturing dynamic user interaction patterns. However, many deep learning-based models struggle to balance short-term interests with long-term preferences and often fail to fully exploit temporal information within interaction sequences. This paper proposes the MTSR-FLP model, which utilizes an MLP-based state generation module to simultaneously consider long-term and short-term preferences. The framework integrates a global representation learning approach to capture general user trends and a local representation learning approach for historical user-specific information. Furthermore, a novel multi-temporal embedding scheme is developed using multiple kernels for absolute and relative timestamps. Experimental results on five real-world datasets demonstrate that MTSR-FLP significantly outperforms advanced baselines, improving HR@10 by up to 31.10% and NDCG@10 by up to 42.54%.
Introduction
Sequential recommendation (SR) is the engine behind personalized content delivery, predicting a user's next interaction based on their historical sequence. Despite the success of deep learning in this domain, many models suffer from "recency bias," over-prioritizing the most recent clicks while neglecting broader user habits. Additionally, time is often treated as a simple ordinal index rather than a rich contextual signal. MTSR-FLP addresses these gaps by treating time as a multi-dimensional feature and fusing global behavioral trends with local user history.
Methodology
The MTSR-FLP architecture is built upon three core components: multi-temporal embedding, dual-track representation learning (local and global), and an item similarity gating mechanism.
Multi-temporal Embedding Learning
Rather than a single time vector, MTSR-FLP employs a sophisticated embedding scheme that captures both absolute and relative time patterns.
- Absolute Embeddings: Combine sequence indices with time-of-day information.
- Relative Embeddings: Calculate intervals between interactions using sine, exponential, and logarithmic kernels to capture patterns ranging from immediate bursts to long-term periodic habits.
Local and Global Representation
- Local Representation Learning (LRL): This module uses self-attention to weigh a user's historical interactions, enriched by the multi-temporal embeddings, to capture specific personal preferences.
- Global Representation Learning (GRL): Utilizing a Location-Based Attention (LBA) layer, this module identifies general item-to-item relationships across the entire user base, providing a "global" context that balances idiosyncratic local patterns.
Item Similarity Gating
A final gating mechanism coordinates the output. It calculates the similarity between a candidate item and the user's recent history to dynamically decide the weight of local versus global preferences for a specific recommendation.
Experimental Results
The researchers evaluated MTSR-FLP against state-of-the-art models like SASRec and GRU4Rec on five diverse datasets: Tmall, Steam, Amazon Games, Amazon Beauty, and Foursquare.
- Performance Gains: MTSR-FLP consistently outperformed baselines. Compared to SASRec, the model improved HR@10 by 6.68% to 31.10% and NDCG@10 by 8.60% to 42.54%.
- Domain Effectiveness: The most dramatic improvements occurred in the Steam dataset, suggesting the model is exceptionally effective in domains with strong temporal dependencies, such as gaming.
- Ablation Insights: Ablation studies confirmed that the multi-temporal approach was the primary driver of success; models utilizing only a single type of temporal embedding showed significantly lower performance.
Conclusion
MTSR-FLP demonstrates that integrating multi-dimensional temporal context with fused local and global preferences significantly enhances sequential recommendation accuracy. Future work may involve incorporating Graph Neural Networks (GNNs) to capture higher-order relationships and expanding the feature set to include item categories and user demographics.
Summary updated