Source-linked AI summary
Sequence-Aware Recommender Systems
Massimo Quadrana, Paolo Cremonesi, Dietmar Jannach
TL;DR
Sequence-aware recommendation addresses limits of matrix-completion models by using sequential interaction logs to capture short-term interests and longer-term patterns. This review categorizes recommendation tasks and goals, surveys algorithms, discusses benchmarking, and identifies open challenges.
Problem
Matrix-completion models generally omit users’ short-term behavior, intents, and sequentially ordered interaction logs available in practical applications.
Method
The review categorizes sequence-aware recommendation scenarios and tasks, surveys algorithms for extracting and leveraging interaction patterns, and formalizes recommendations as ordered item lists.
Results
The review identifies a scattered research field lacking common understanding of its problem facets and evaluates existing approaches across scenarios, algorithms, and benchmarking issues.
Takeaways & Limitations
The review aims to support more standardized and reproducible research on sequence-aware recommender systems.
Takeaways & Limitations
Varied evaluation protocols and metrics, with research still largely focused on accuracy, make progress and recommendation usefulness difficult to assess.
Abstract
from arXiv · showhide
Recommender systems are one of the most successful applications of data mining and machine learning technology in practice. Academic research in the field is historically often based on the matrix completion problem formulation, where for each user-item-pair only one interaction (e.g., a rating) is considered. In many application domains, however, multiple user-item interactions of different types can be recorded over time. And, a number of recent works have shown that this information can be used to build richer individual user models and to discover additional behavioral patterns that can be leveraged in the recommendation process. In this work we review existing works that consider information from such sequentially-ordered user- item interaction logs in the recommendation process. Based on this review, we propose a categorization of the corresponding recommendation tasks and goals, summarize existing algorithmic solutions, discuss methodological approaches when benchmarking what we call sequence-aware recommender systems, and outline open challenges in the area.
1 INTRODUCTION
Sequence-aware recommendation extends the dominant matrix-completion formulation by using ordered, heterogeneous interaction logs to model short-term intent and longer-term behavioral patterns. This survey categorizes the resulting tasks, reviews algorithms, discusses benchmarking, and identifies open challenges.
- Matrix completion predicts missing user-item ratings but typically does not model short-term behavior, intent, or sequential interaction logs.
- Session-based recommendation adapts to anonymous users’ short-term interests by matching recommendations to a sequence of observed actions.
- Sequential logs support next-item prediction, interest-drift detection, popularity-trend identification, and timing reminders for previously seen or purchased items.
- The literature is scattered, so the survey categorizes scenarios, reviews algorithms, and discusses benchmarking to support more standardized and reproducible research.
- The paper covers problem characterization, algorithmic approaches, evaluation methodology, and future research directions across Sections 2–6.
2 CHARACTERIZING SEQUENCE-AWARE RECOMMENDER SYSTEMS
Sequence-aware recommenders use enriched, ordered action logs to produce ordered recommendation lists rather than isolated utility predictions. Their tasks include learning sequence-related patterns, handling order and timing constraints, and optimizing list-level utility.
- Inputs: The main input is an ordered, often timestamped list of user actions, which may come from known or anonymous users.
- Inputs: Inputs can be enriched clickstream data containing user, item, action-type, and additional attributes, while explicit ratings are generally not assumed.
- Outputs: The output is an ordered list of items, and the recommendation order can itself matter in scenarios where users should consider items sequentially.
- Computational Tasks: Common computational tasks identify sequential, co-occurrence, or distance patterns in user actions, either explicitly or implicitly within machine-learning models.
- Computational Tasks: Sequence-aware recommenders may reason over strict, heuristic, or implicit order constraints and relate learned patterns to the time at which recommendations are requested.
- Abstract Characterization: Formally, the system computes an ordered list of length k from possible item-list permutations, maximizing a learned utility function for each user.
- Abstract Characterization: Unlike item-level matrix completion, the utility function can score whole ordered lists for diversity, transitions, and satisfaction of order constraints.
3 A CATEGORIZATION OF SEQUENCE-AWARE RECOMMENDATION TASKS
The survey organizes sequence-aware recommendation into four overlapping goals: context adaptation, trend detection, repeated recommendation, and order or sequential-pattern handling. It further distinguishes context settings by available history and emphasizes that intent and evaluation remain open challenges.
- The literature’s four main goals are context adaptation, trend detection, repeated recommendation, and consideration of order constraints and sequential patterns.
- These categories share one formal problem characterization, may require different sequence-aware algorithms, and can be combined in a single application.
- Context Adaptation: Interactional context, such as shopping intent or mood, is not directly observable and must be inferred from recent actions and behavioral patterns.
- Context Adaptation: Context-adaptation settings vary according to historical-user data and the importance assigned to recent interactions.
- Context Adaptation: The taxonomy includes last-N recommendation, session-based recommendation, and session-aware recommendation using both current-session and prior behavior.
- Context Adaptation: Combining older interactions for long-term modeling with recent interactions for short-term intent can yield recommendations mixing familiar and novel items.
- Context Adaptation: Many context-adaptation studies implicitly define recommendation quality as predicting hidden session actions rather than specifying the application’s utility or purpose.
3.2 Trend Detection
Sequence-aware recommenders can detect trends in sequential interaction data, although trend detection is less explored. These trends may reflect changing popularity across communities or changing interests of individuals.
- Trend detection is a less explored goal for sequence-aware recommenders.
- Community trends capture long-term, seasonal, or short-term changes in item popularity over time.Recent popularity in a domain such as fashion may inform recommendation selection.
- Individual trends capture users’ changing interests caused by natural preference drift, social influence, exceptional events, or discovery.The passage gives changing news interests as an application example.
3.3 Repeated Recommendation
Repeated recommendation scenarios involve recommending items that users already know or previously purchased, which traditional matrix completion does not address. Sequence-aware systems can identify repeated behavior and must also determine when recommendations should be made.
- Repeated recommendation scenarios include items a user already knows or purchased, outside traditional matrix completion.
- Past interaction logs can reveal repeated behavior patterns at the individual-user or community level.Repeated purchases of consumables and repeated app use are application examples.
- Repeated recommendation requires reasoning about timing in addition to selecting items.Reminder timing may be narrow when items can become obsolete, while frequent reminders may be inappropriate.
3.4 Consideration of Order Constraints and Observed Sequential Patterns
Sequence-aware recommendation can use both externally specified ordering constraints and sequential patterns inferred from user behavior. Systems may also need to determine whether recommendation order matters and ensure applicable constraints are respected.
- Consideration of Order Constraints and Observed Sequential Patterns: Ordering can matter in sequence-aware tasks beyond the user’s last session.
- Consideration of Order Constraints and Observed Sequential Patterns: Recommendation ordering can incorporate two information sources: external domain constraints and observed user-consumption sequences.
- Consideration of Order Constraints and Observed Sequential Patterns: External constraints may require or suggest course prerequisites or viewing a movie sequel after its preceding episode.
- Consideration of Order Constraints and Observed Sequential Patterns: Sequential pattern mining can infer consumption relationships for tasks such as predicting website navigation or selecting next music tracks.
- Consideration of Order Constraints and Observed Sequential Patterns: The importance of recommendation order depends on the application’s goals, and resulting lists must respect applicable strict or weak constraints.
3.5 Categorization of Existing Works
The survey categorizes existing sequence-aware recommendation works by context adaptation, ordering constraints, application domain, and specific tasks, revealing uneven coverage across these dimensions.
- Categorization scheme: The categorization classifies works by context-adaptation type, ordering constraints, and application domain.The survey uses these dimensions to assess which aspects are well explored and which need further investigation.
- Context adaptation: A large fraction of reviewed papers uses only the last visited object, while a slightly smaller group addresses session-based recommendation.More than half of session-based and session-aware papers were published from 2014 onward.
- Context adaptation: Even fewer works combine sequence-aware recommendation with users’ long-term preferences, although some studies report that this can improve recommendation accuracy.The survey identifies the interplay between short-term intents and long-term preferences as requiring further research.
- Ordering constraints: Most reviewed works rely on implicitly derived weak ordering constraints, whereas only a few address explicit strong constraints.The survey links this imbalance partly to research concentration in media recommendation, where ordering constraints may be present but often remain unused algorithmically.
- Application domains: E-commerce is the most investigated application domain, alongside music, points of interest, web navigation, app usage, and several less explored domains.Examples of less explored areas include advertisements, news, jobs, video streaming, query recommendation, and workflow activities.
- Specific tasks: The survey identifies repeated recommendation and trend detection as relatively uncommon specific tasks, while next-item recommendation and list continuation are frequent goals.Next-basket recommendation is also reported as a focus of only a few works.
4 ALGORITHMS FOR SEQUENCE-AWARE RECOMMENDER SYSTEMS
The survey organizes algorithms for extracting patterns from sequential user-action logs into three main classes. These are sequence learning, sequence-aware matrix factorization, and hybrid methods, with sequence learning used by most reviewed works.
- Algorithm taxonomy: Three main algorithm classes are sequence learning, sequence-aware matrix factorization, and hybrids.The taxonomy further divides these classes into subcategories and lists uncommon technical approaches under “Others”.
- Sequence learning: Most reviewed works rely on sequence learning methods to model patterns in sequential interaction logs.The survey describes sequence learning as a natural fit for predicting future user actions from past observations.
4.1 Sequence Learning
Sequence learning methods model ordered user actions to predict future recommendations, using approaches ranging from frequent patterns and Markov models to recurrent neural networks.
- Frequent Pattern Mining: Frequent pattern methods mine ordered consumption patterns offline and use association rules or compact knowledge representations during prediction.Rules typically carry strength values such as confidence and support.
- Application Examples: Personalized pattern-relevance scoring schemes yield more accurate personalized next-item recommendations than a popularity-based method.
- Sequence Modeling Categories: Sequence modeling methods for recommendation mainly comprise Markov Models, Reinforcement Learning, and Recurrent Neural Networks.
- Markov Models: Markov-based recommenders restrict dependencies to finite histories, while higher-order models capture longer relationships but face sparsity and order-selection challenges.Heuristics, mixtures, variable-order models, and hidden states are used to address these issues.
- Recurrent Neural Networks: Recurrent neural networks update a hidden state from each input and the previous state to predict probabilities for subsequent items.
4.2 Sequence-aware Matrix Factorization
Sequence-aware matrix factorization incorporates order or timing into factorization-based recommendation, retaining standard algorithms while introducing encoding and adaptation challenges.
- Problem Formulation: Sequence-aware matrix-factorization approaches extend matrix completion with sequence information, including time-derived ordering, for problems such as repeated-recommendation timing.
- Application Examples: Applications include completing prefix-rating matrices for interactive story continuations and modeling session events and items for next-in-sequence e-commerce recommendation.
- Discussion: A key advantage is that standard matrix-factorization algorithms can often be applied, but application-specific encodings must remain suitable and computationally feasible.Some encodings require feedback at each sequence step and can create huge matrices through combinatorial growth.
- Time-aware Extensions: Time-aware factorization methods generally track behavior changes over large time spans but often cannot be updated in real time.They may therefore need extension or combination with methods supporting short-term adaptation.
4.3 Hybrid methods
Hybrid methods combine sequence learning with factorization, topic models, or other predictors to integrate short- and long-term preferences and address individual method shortcomings.
- Hybrid Model Rationale: Hybrid models combine sequence-learning flexibility with factorization-based robustness to data sparsity and can incorporate collaborative learning-to-rank models such as BPR.
- Unified Models: FPMC combines matrix factorization with Markov Chains to recommend the next item from a user’s last-N interactions.The method represents first-order interactions as a three-dimensional user-current-item-next-item tensor.
- Hybrid Variants: Other hybrids combine Markov Chains with item similarity, LDA, or behavioral clustering across unified, cascade, and meta-level designs.
- Discussion: Hybrid approaches are used to address shortcomings such as limited content discovery and user- or item-cold-start situations.Design challenges include combining prediction channels and determining their importance weights.
4.4 Other methods
Other sequence-aware recommenders use graphs or discrete optimization to represent evolving preferences, order constraints, contextual behavior, and structured sequence-generation objectives.
- Graph-based Methods: Graph-based methods fuse long-term and short-term preferences, model evolving interests, or represent sequential patterns through transition graphs.
- Graph-based Methods: Graph representations also support new-user news recommendation by organizing browsing behavior across sessions and referrer domains.
- Discrete Optimization Methods: Discrete optimization addresses weak or strict ordering constraints in travel planning, course sequencing, and playlist generation.
- Application Examples: Playlist and course-sequence recommenders use scoring followed by greedy reranking or optimize graduation time, GPA, interests, and prerequisite constraints.
- Discussion: Discrete optimization methods generally use heuristic search or greedy optimization instead of exact exhaustive search because of underlying computational complexity.
4.5 Summary and Pros and Cons of Selected Approaches
The survey summarizes major algorithm families for sequence-aware recommendation, presenting their core ideas alongside typical advantages and disadvantages. These entries provide rough orientation rather than exhaustive characterizations of every algorithm.
- The listed pros and cons are intended as a rough orientation across algorithm families.
- Specific algorithms within a family may have additional advantages or disadvantages beyond the table’s general entries.
- Table 4 organizes selected sequence-aware recommendation algorithms by their main ideas, advantages, and disadvantages.
5 EVALUATION OF SEQUENCE-AWARE RECOMMENDER SYSTEMS
Evaluation of sequence-aware recommenders extends beyond traditional offline matrix-completion testing because temporal order, sessions, event types, and future actions shape both protocols and targets. The survey therefore compares partitioning, prediction, validation, and metric choices, while highlighting unresolved standards for evaluating recommendation lists.
- Offline, laboratory, and field studies can evaluate sequence-aware recommenders, but field tests remain comparatively rare in academic research.
- Dataset partitioning: Training-test data can be split at event or session level, with partitioning applied across the community or separately for users.
- Dataset partitioning: No common partitioning standard exists; session-level splitting is advisable for session-based or session-aware evaluation, while user-level partitioning supports new-user scenarios.
- Given-N next-item prediction variants reveal immediate future actions, optionally using look-ahead or earlier interactions through look-back.
- Evaluation methodology should be chosen according to the research question and application domain.
- Recommendation-list diversity and coherence lack standard metrics whose alignment with users’ quality perceptions is established.
6 SUMMARY AND FUTURE DIRECTIONS
The survey frames sequence-aware recommendation as practically important but identifies unresolved challenges in intent detection, profile integration, data use, and evaluation. Future progress requires methods and benchmarks that better reflect changing context, richer interactions, and users’ goals.
- Sequence-aware recommendation is practically relevant, especially in session-based and session-aware settings, but the field still contains open research questions.
- Intent detection: Short-session interactions may not reliably reveal whether users seek discovery, known content, browsing, or purchase-oriented actions.
- Combining short-term and longer-term profiles: Better integrated models are needed to combine predominant short-term intentions with longer-term behavioral patterns and preferences.
- Leveraging additional data and general trends: Research should use richer interaction and referral data while detecting short-term trends and interest shifts across user communities.
- Toward standardized and more comprehensive evaluations: Evaluation remains fragmented and accuracy-focused, motivating standardized, purpose-oriented approaches that account for users’ contextual situations and goals.