Abstract The emerging topic of sequential recommender systems (SRSs) has attracted increasing attention in recent years. Different from the conventional recommender systems (RSs) including collaborative filtering and content-based filtering, SRSs try to understand and model the sequential user behaviors, the interactions between users and items, and the evolution of users’ preferences and item popularity over time. SRSs involve the above aspects for more precise characterization of user contexts, intent and goals, and item consumption trend, leading to more accurate, customized and dynamic recommendations. In this paper, we provide a systematic review on SRSs. We first present the characteristics of SRSs, and then summarize and categorize the key challenges in this research area, followed by the corresponding research progress consisting of the most recent and representative developments on this topic. Finally, we discuss the important research directions in this vibrant area.
Introduction
Sequential Recommender Systems (SRSs) represent a significant evolution in recommendation technology, moving beyond the static paradigms of traditional collaborative filtering and content-based filtering. While traditional systems are adept at capturing long-term user preferences, they often overlook the critical temporal context and the specific order of user behaviors. In real-world applications, user actions are rarely isolated; they are part of a dynamic sequence where current choices are heavily influenced by recent history and evolving trends.
The Sequential Recommendation Problem
The fundamental goal of an SRS is to predict a user's next action by analyzing their historical sequence of interactions. Unlike static models that treat user-item interactions as a fixed set, SRSs model these interactions as a chronological stream. This approach allows for a more precise characterization of user intent and the dynamic nature of item popularity.
As shown in the figure above, a user like Jimmy might follow a logical sequence of travel-related actions (flight booking followed by a hotel reservation and car rental), while Tina's sequence reflects a progression of tech-related purchases. Capturing these dependencies is essential for providing accurate, context-aware recommendations.
Key Challenges in SRS Research
Developing effective SRSs requires addressing several unique data characteristics and modeling hurdles:
- Long-term Dependencies: Capturing relationships across extended sequences remains a primary challenge for many architectures.
- Flexible Ordering: In many domains, the order of actions is not strictly fixed. For example, a user might buy a phone case before or after a screen protector, but both actions are collectively dependent on the initial purchase of a phone. Models must be flexible enough to handle this "fuzzy" ordering.
- Data Noise: Real-world interaction data is inherently messy. Accidental clicks or irrelevant browsing sessions can introduce noise that obscures the user's true underlying intent. Models that fail to account for this may learn false dependencies.
- Heterogeneous Relations and Hierarchies: User sequences often exist within complex hierarchies, where demographic factors influence session-level intents, which in turn dictate specific item interactions.
Methodological Taxonomy
The field has progressed from simple statistical models to complex deep learning architectures:
1. Traditional Sequence Models
Early approaches utilized sequential pattern mining and Markov Chain models. While intuitive, Markov chains are limited by the "Markov property," which assumes the next state depends only on the immediate previous state, thus failing to capture long-term dependencies or collective behaviors.
2. Latent Representation Models
Techniques such as Factorization Machines and basic embeddings map users and items into a latent space. These models calculate interaction scores based on distances within this space, offering a more nuanced representation than simple pattern matching.
3. Deep Neural Network Models
Modern SRSs leverage advanced deep learning to overcome the limitations of earlier models:
- Recurrent Neural Networks (RNNs): LSTMs and GRUs are widely used for modeling long-term dependencies, though they can struggle with flexible item ordering due to their rigid chronological assumptions.
- Convolutional Neural Networks (CNNs): By treating interaction embeddings as a 2D matrix, CNNs can extract local patterns without assuming a strict point-to-point order, making them effective for handling flexible ordering.
- Graph Neural Networks (GNNs): These models map sequences to paths on directed graphs, capturing complex, non-linear transitions between items.
- Attention and Memory Mechanisms: Attention mechanisms help models focus on relevant historical interactions while down-weighting noise. Memory networks provide explicit storage to maintain and update long-term user history.
Future Research Directions
Several promising frontiers remain for the SRS community:
- Context-aware SRSs: Integrating external factors such as time of day, location, and device type.
- Social-aware SRSs: Incorporating social influence and peer behaviors into sequential modeling.
- Interactive SRSs: Modeling the continuous, multi-step feedback loop between users and platforms.
- Cross-domain SRSs: Leveraging sequential footprints across different domains to enhance prediction accuracy.
By addressing these challenges, Sequential Recommender Systems will continue to provide more personalized, dynamic, and contextually relevant user experiences.
Summary updated