Sequential Recommender Systems: Challenges, Progress and Prospects

Motivation: Why Sequential Recommender Systems?

• SRSs model sequential dependencies to capture users’ current and recent preferences, unlike traditional content-based and collaborative filtering systems, which model interactions statically [27].The paper notes that SRS interactions can include views or purchases.Source: §1 · p. 1
• Successive actions depend on prior actions, a common pattern in transaction data that conventional content-based and collaborative filtering systems cannot capture well [12].The paper illustrates this with travel bookings and purchases preceding a possible next action.Source: §1 · p. 1
• Because user preferences and item popularity change over time, SRSs can capture dynamics that matter for accurately profiling users and items.The paper gives changing phone preferences and iPhone popularity as examples.Source: §1 · p. 1
• Using prior interactions as context helps an SRS predict near-future choices and diversify recommendations by avoiding items identical or similar to those already chosen.Source: §1 · p. 2

Formalization: What are Sequential Recommender Systems?

• An SRS ranks candidate items to predict future interactions from a sequence whose elements can encode a user, action, and item, alongside metadata, action types, and context.The ranking score may use a conditional probability or interaction score, and the output is an ordered list.Source: §1 · p. 2

• SRS learning is challenging because interaction sequences have complex elements such as triples, motivating a systematic review of their challenges and progress.Source: §1 · p. 2

Contributions

• The paper categorizes data-driven SRS challenges, surveys state-of-the-art technical progress, and discusses prospects for the community.Source: §1 · p. 2

2.1 Handling Long User-Item Interaction Sequences

• Long sequences make cascading higher-order and distant dependencies difficult: higher-order Markov models [6] and RNNs have limits, while LSTM/GRU RNNs risk false links and mixtures [15] remain scarce.Higher-order Markov models are limited by parameter growth, and RNNs by strong order assumptions; mixture models combine temporal ranges to capture short- and long-term dependencies.Source: §2 · pp. 2–3

2.2 Handling User-Item Interaction Sequences with a Flexible Order

• Flexible-order sequences require collective rather than point-wise dependency modeling; Markov-chain, factorization-machine, and RNN SRSs miss collective relations, and only a few CNN approaches [17] [26] address them.For example, milk and butter need not be ordered relative to each other, while their combination can predict flour.Source: §2 · p. 3

2.3 Handling User-Item Interaction Sequences with Noise

• Noisy sequences mix relevant, weakly relevant, and irrelevant interactions, so SRSs must learn dependencies attentively and discriminatively; attention models and memory networks selectively use relevant interactions.In the example sequence, the rose is noise, while the next milk purchase depends on bacon, eggs, and bread.Source: §2 · p. 3

2.4 Handling User-Item Interaction Sequences with Heterogeneous Relations

• Heterogeneous sequences combine relations such as item similarity and short- versus long-term dependencies, which must be modeled distinctly and collaboratively; mixture models [12] [15] are the only reported solution.Mixture models combine sub-models that represent different relation types.Source: §2 · p. 3

2.5 Handling User-Item Interaction Sequences with Hierarchical Structures

• Hierarchical dependencies arise from user/item metadata and historical sessions; feature-enriched models address item features, while demographics remain rarely considered and hierarchical embeddings, RNNs [13], and attention networks incorporate prior sessions.Historical subsequences may influence the next interaction within the current session.Source: §2 · pp. 3–4

3.1 Traditional Sequence Models for SRSs

• Traditional sequence models use pattern mining or Markov transitions, but pattern mining loses infrequent items and Markov models miss long-term and collective dependencies.Markov approaches calculate transitions directly from observations [3] or embed chains in Euclidean space; few pattern-mining works are reported, including.Source: §3.1 · pp. 4–5

3.2 Latent Representation Models for SRSs

• Latent representation models capture complex dependencies in latent space, but factorization methods are vulnerable to data sparsity, while embedding-based approaches offer simplicity, efficiency, and efficacy.Factorization models factorize interaction data into user and item latent factors; embedding approaches use representations to score interactions or successive actions or distances.Source: §3.2 · p. 5

3.3 Deep Neural Network Models for SRSs

• RNNs dominate deep-learning SRS research but can impose false adjacent dependencies and miss collective effects, while CNNs relax order assumptions yet struggle with long-term dependencies.LSTM and GRU variants aim to capture long-term dependencies, and hierarchical RNNs model more complex sequence dependencies [13]. CNNs learn sequential patterns as local features from interaction embeddings [17] [26].Source: §3.3 · p. 6
• GNN-based SRSs construct graphs from interaction sequences to capture complex relations and may enable explainable recommendations, but remain at an early stage.Each interaction becomes a graph node and each sequence a path; graph embeddings capture relations across the structure [22].Source: §3.3 · p. 6
• Attention filters irrelevant interactions, memory networks store histories or item knowledge, and mixture models combine specialized dependency-capturing models [15].Memory networks can reduce interference from irrelevant interactions, but memory and mixture models remain insufficiently studied or early-stage.Source: §3.3 · p. 6

Context-aware sequential recommender systems

• Context-aware SRSs should account for user or item context that changes over time, an important factor ignored by most existing systems.The paper identifies this as an important direction for future work.Source: §4 · pp. 6–7

Social-aware sequential recommender systems

• Social-aware SRSs should incorporate the influence of other people’s behaviors and opinions on users’ choices, which existing work usually ignores.Source: §4 · p. 7

Interactive sequential recommender systems

• Interactive SRSs should model continuous user-platform interactions and generate recommendations across multiple time steps rather than for one action at a time.The paper highlights user-seller interactions as a promising direction.Source: §4 · p. 7

Cross-domain sequential recommender systems

• Cross-domain SRSs should leverage sequential dependencies between domains, such as buying car insurance after purchasing a car, for more accurate and diverse recommendations.Most existing SRSs ignore these cross-domain dependencies.Source: §4 · p. 7

5 Conclusions

• SRSs have been at the core of the recommender-systems field in the past three to five years, offering recommendations intended to satisfy daily requirements.The passage describes these recommendations as more intelligent and favorable.Source: §5 · p. 7
• Recommender systems are described as one of the most direct and practical applications of artificial intelligence in daily life.Source: §5 · p. 7
• The authors hope this summary gives the recommender-systems research community an overview of SRS challenges, recent progress, and future directions.Source: §5 · p. 7