PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey on Session-based Recommender Systems

Shoujin Wang, Longbing Cao, Yan Wang, Quan Z. Sheng, Mehmet Orgun, Defu Lian

arXiv:1902.04864v3cs.IR

TL;DR

Existing recommender systems often emphasize long-term, static preferences, while session-based recommender systems target short-term, dynamic preferences; unified problem statements and a consolidated research landscape were lacking. This paper reviews SBRSs systematically, proposes unified problem and classification frameworks, analyzes session-data challenges, and discusses open issues and research directions.

  • Problem

    Unified problem statements, detailed accounts of session-data characteristics and challenges, and a clear overview of SBRS research progress were lacking.

  • Method

    The paper conducts a systematic review, formalizes SBRS concepts and problems, and classifies existing approaches by their session-data dependencies and research sub-areas.

  • Results

    The review organizes SBRS research into unified frameworks, analyzes session-data characteristics and challenges, and compares representative approach classes.

  • Takeaways & Limitations

    The paper identifies open issues and promising directions for advancing session-based recommendation, including richer contextual and cross-domain information.

  • Takeaways & Limitations

    Raw POI data lacks session structure, so researchers may need to construct sessions manually by grouping a user’s check-ins within a single day.

Abstract

from arXiv · show

Recommender systems (RSs) have been playing an increasingly important role for informed consumption, services, and decision-making in the overloaded information era and digitized economy. In recent years, session-based recommender systems (SBRSs) have emerged as a new paradigm of RSs. Different from other RSs such as content-based RSs and collaborative filtering-based RSs which usually model long-term yet static user preferences, SBRSs aim to capture short-term but dynamic user preferences to provide more timely and accurate recommendations sensitive to the evolution of their session contexts. Although SBRSs have been intensively studied, neither unified problem statements for SBRSs nor in-depth elaboration of SBRS characteristics and challenges are available. It is also unclear to what extent SBRS challenges have been addressed and what the overall research landscape of SBRSs is. This comprehensive review of SBRSs addresses the above aspects by exploring in depth the SBRS entities (e.g., sessions), behaviours (e.g., users' clicks on items) and their properties (e.g., session length). We propose a general problem statement of SBRSs, summarize the diversified data characteristics and challenges of SBRSs, and define a taxonomy to categorize the representative SBRS research. Finally, we discuss new research opportunities in this exciting and vibrant area.

1 INTRODUCTION

The paper introduces SBRSs to model short-term, dynamic preferences from session data, addressing fragmented definitions and the lack of a unified account of the field.

  • Traditional recommender systems often learn long-term, static preferences from all historical interactions, whereas current choices also reflect recent preferences and time-sensitive context.
  • SBRSs use sessions as the basic input unit to capture short-term preferences and preference dynamics for timely recommendations.A session contains multiple user-item interactions occurring within a continuous period, such as one transaction visit.
  • The literature contains diverse SBRS terms, settings, assumptions, and application domains, producing inconsistencies and no unified problem statement or systematic challenge analysis.
  • The paper formalizes SBRSs through the concepts of user, item, action, interaction, and session, while covering recommendations for current-session interactions or future sessions.
  • This survey responds with a comprehensive overview of session-data characteristics and SBRS challenges, a systematic classification and comparison of approaches, technical summaries, and open research directions.

2 RELATED WORK

The related-work discussion distinguishes SBRSs from sequence-based recommendation and organizes varied SBRS tasks and prior reviews within a unified framework.

  • SBRSs vs. SRSs: SBRSs use session data with explicit boundaries, whereas SRSs use ordered sequence data that typically lacks multiple session intervals and boundaries.
  • SBRSs vs. SRSs: SBRSs may predict an unknown part of the current session or a future session by learning intra- or inter-session dependencies, which can be sequential or non-sequential.
  • SBRS sub-areas: The unified task framework separates next-interaction, next partial-session, and next-session recommendation according to the recommendation target.
  • Related surveys: Earlier reviews and empirical studies covered selected SBRS algorithms or mainly ordered sessions, leaving the field without an extensive review of its characteristics, challenges, progress, and open issues.
  • Related surveys: The survey aims to bridge these gaps by comprehensively reviewing SBRS research and supporting further development of the area.

3 SBRS PROBLEM STATEMENT

The paper formalizes SBRSs around users, items, actions, interactions, and sessions, then specifies their inputs and recommendation outputs. It further describes session properties that shape the SBRS problem.

  • Formal entities: An SBRS is defined using formal concepts of users, items, actions, interactions, and sessions.These entities and behaviours form the basis for characterizing and categorizing SBRSs.
  • Users and items: A user takes actions on items and receives recommendation results, while user information may be unavailable, making a session anonymous.User attributes can affect actions and sessions, but privacy protection or absent login information may prevent user identification.
  • Users and items: An item is an entity to be recommended, such as a product or service, with identifying and descriptive attributes.Item types vary across domains, including news articles, products, and courses.
  • Interactions and sessions: An interaction is a user-item-action tuple, with anonymous or item-only forms available when user information or action diversity is absent.All interactions together form the interaction set O.
  • Interactions and sessions: A session is a non-empty bounded list of interactions generated during continuous time, potentially linked to user or session information.Sessions may contain duplicated interactions and attributes such as duration, time, and day.
  • Input and output: SBRS inputs include known current-session interactions, historical sessions, or both, while outputs predict subsequent interactions or a future session.Inputs may be anonymous or non-anonymous, ordered or unordered, and contain single or multiple action types.

4 CHARACTERISTICS AND CHALLENGES

Session data vary in length, internal order, action types, user information, and hierarchical structure. These characteristics create distinct challenges for modelling dependencies and producing recommendations.

  • Session length: Session length divides sessions into long, medium, and short types, with dataset-specific definitions.Long sessions may contain more context but also irrelevant interactions, whereas short sessions provide limited recommendation information.
  • Session length: Medium sessions commonly contain 4 to 9 interactions and usually provide necessary context without many irrelevant interactions.The paper reports medium sessions as the most common case in transaction-record data from e-commerce.
  • Internal order: Sessions may be unordered, strictly ordered, or flexibly ordered, requiring models to handle co-occurrence, sequential, or mixed dependencies.Flexibly ordered sessions combine sequentially dependent interactions with randomly inserted interactions.
  • Action type: Multi-type-action sessions contain complex dependencies within and across action types, unlike single-type-action sessions with one dependency type.Online shopping sessions may combine clicks for comparison with subsequent purchases.
  • User information: Non-anonymous sessions connect sessions from the same user, whereas anonymous sessions generally provide only current-session context.The resulting challenge is modelling evolving long-term preferences or personalized preferences from limited information.
  • Session-data structure: Single-level data expose only within-session dependencies, while multi-level data require modelling dependencies within and across hierarchical levels.Multi-level data may include attribute and session levels in addition to the interaction level.

5 CLASSIFICATION AND COMPARISON OF SBRS APPROACHES

The survey classifies SBRS approaches by technique into three super-classes and compares them by learnable dependencies, applicability, and research activity. Conventional, latent-representation, and deep-neural approaches offer different trade-offs in complexity, computation, and modeling capacity.

  • Classification: The taxonomy identifies conventional, latent representation, and deep neural network approaches as three SBRS super-classes.These contain eight classes and nine deep-learning sub-classes, yielding 15 atomic classes overall.
  • Comparison: Conventional approaches are simple and interpretable, latent representations balance computation and representation learning, and deep networks target complex dependencies with greater computational demands.The comparison frames the approaches as differing in complexity, computation cost, and ability to model complex or sparse data.
  • Comparison: The survey compares the 15 atomic classes according to the types of dependencies they can learn from session data.Examples include sequential, short-term, first-order, and pointwise dependencies for Markov-chain approaches.
  • Research landscape: The survey manually counted SBRS publications by atomic class using Google Scholar searches conducted on 20 March 2021.Searches used session and recommendation-related keywords, followed by manual relevance filtering and counting.

6 CONVENTIONAL SBRS APPROACHES

Conventional SBRS approaches apply pattern mining, KNN, Markov chains, and related probabilistic techniques to model dependencies in session data. They are generally simple and effective, with session-KNN showing strong empirical accuracy in several comparisons.

  • Pattern/rule-based approaches: Pattern-mining SBRSs use frequent or association patterns for unordered sessions and sequential patterns for ordered sessions.Their common pipeline consists of mining patterns, matching the current session or sequence, and generating recommendations.
  • Extensions: Conventional approaches are simple, straightforward, and easy to implement, while variants incorporate weighting, higher-order transitions, hidden states, factorization, or hybrid models.These extensions address differences in page significance, transition order, latent transitions, personalization, and combinations with collaborative filtering or other techniques.
  • KNN-based approaches: KNN-based SBRSs retrieve similar interactions or sessions and score candidate items according to their similarity to the current context.Session-KNN uses the whole session context, whereas item-KNN focuses on the current item and co-occurrence vectors.
  • Markov chain based approaches: Markov-chain SBRSs model transitions among interactions or sessions to predict probable next interactions or sessions.Basic models calculate transition probabilities, predict paths, match the current context to high-probability paths, and recommend subsequent items.
  • Comparison: KNN-based approaches, especially session-KNN, achieved superior prediction accuracy in an empirical comparison across seven retail, music, and news datasets.The comparison covered the first three conventional approach classes.

7 LATENT REPRESENTATION APPROACHES FOR SBRSs

Latent representation SBRSs encode session interactions in low-dimensional spaces using shallow factorization or neural models. These representations capture dependencies for recommendation while offering lower complexity than deep neural approaches.

  • Overview: Latent representation approaches learn low-dimensional representations of interactions with shallow models and use them for session-based recommendations.They are divided into latent factor model and distributed representation approaches.
  • Latent factor model based SBRSs: Latent factor models factorize observed transition structures to estimate unobserved transitions for recommendation.The transition tensor records user-conditioned probabilities from one interaction to another, and factorization produces latent representations.
  • Latent factor model based SBRSs: The factorized personalized Markov-chain formulation models predicted transitions through interactions among user, last-interaction, and current-interaction latent vectors.A pairwise formulation is introduced to alleviate the negative effects of sparse observed transitions.
  • Distributed representation based SBRSs: Distributed-representation approaches use shallow neural structures similar to Skip-gram or CBOW to map interactions into a low-dimensional latent space.The learned representations encode intra- or inter-session dependencies, with user and item embeddings optionally incorporated.
  • Extensions: Variants address cold-start items, interaction relevance, and next-basket recommendation through compound item representations, attention, and hierarchical representations.These extensions alter how item features, session relevance, or basket structure are represented.
  • Comparison: Latent representation approaches are more complex than conventional methods but less complex and less computationally costly than deep neural approaches.Their representation learning can nevertheless outperform some conventional and deep neural approaches in reported studies.

8 DEEP NEURAL NETWORK APPROACHES FOR SBRSs

Deep neural network SBRS approaches use deep models to represent complex dependencies within and across sessions. The survey distinguishes basic neural architectures from advanced models.

  • Overview: Deep neural network approaches exploit deep networks to model complex intra- and inter-session dependencies for recommendation.They are divided into basic architectures and advanced models according to their underlying framework.

8.1 Basic Deep Neural Network based SBRSs

Basic deep neural SBRS approaches model session contexts with RNNs, MLPs, CNNs, or GNNs, each suited to different session structures and dependency patterns.

  • Basic deep neural SBRSs comprise RNN-, MLP-, CNN-, and GNN-based approaches.These classes differ in their network architectures and modeling assumptions.
  • RNN-based approaches: RNN-based models represent ordered session contexts as interaction sequences and use the final hidden state to predict the next interaction.GRU4Rec updates hidden states using prior states, candidate states, and gates.
  • RNN-based approaches: Variants enhance RNNs through augmentation, dropout, hierarchical modeling, uncertainty handling, side information, time decay, or attention.Hierarchical RNNs capture both intra- and inter-session dependencies.
  • MLP-based approaches: MLP-based models combine multiple session representations or factors into compound representations and are mainly suited to unordered sessions.Examples combine click and view actions, reminders, popularity, discounts, or long- and short-term preferences.
  • CNN-based approaches: CNN-based SBRSs relax rigid order assumptions while learning local and collective dependencies through filtering and pooling.Max pooling captures the most significant session features for subsequent recommendations.
  • GNN-based approaches: GNN-based SBRSs construct graphs from session interactions and propagate node information to model complex transitions within or between sessions.GGNNs update interaction embeddings from a node’s previous state and neighborhood information; SR-GNN is a pioneering example.

8.2 Advanced Model based SBRSs

Advanced SBRSs augment basic models with attention, memory, mixture, generative, or reinforcement-learning mechanisms to emphasize, retain, combine, generate, or sequentially optimize session information.

  • Advanced SBRSs combine attention, memory networks, mixture models, generative models, or reinforcement learning with basic approaches.These combinations aim to construct more powerful session-based recommenders.
  • Attention models: Attention models weight interactions or sessions by relevance to the target and aggregate them into an informative context representation.The utility function may also use the target interaction, and aggregation is often a weighted sum.
  • Attention models: Hierarchical attention integrates historical and current sessions to capture long- and short-term preferences.Co-attention networks additionally explore correlations between user-related information sources.
  • Memory networks: Memory-network SBRSs maintain interaction embeddings in an external memory matrix and use controller operations to write and read relevant information.During prediction, the target interaction can guide reading from memory, often through attention.
  • Mixture models: Mixture models combine submodels that capture different dependency types, such as short- versus long-term or purpose-specific intra-session dependencies.M3 and MCPRN exemplify this strategy.
  • Generative models: Generative SBRSs generate next interactions or sessions from session context, better matching stepwise shopping behavior.Generation may use probabilistic, utility-based, or stochastic processes.
  • Reinforcement learning: Reinforcement-learning SBRSs formulate user–recommender interaction as an MDP with states, actions, rewards, transitions, and discounting.They map states to weights, score candidate items, and recommend the highest-scoring items.

8.3 Comparison of Deep Neural Network based SBRS Approaches

The survey compares two classes and nine subclasses of deep neural SBRS approaches by scenarios, representative strengths and weaknesses, and typical works.

  • Table 8 compares two deep neural SBRS classes and nine subclasses across applicable scenarios, typical pros and cons, and representative works.

9 SBRS APPLICATIONS, ALGORITHMS AND DATASETS

SBRS applications span conventional product, content, and service recommendations, while algorithms, datasets, and open-source implementations support evaluation across diverse domains.

  • Applications: SBRS applications include conventional next-item recommendations and emerging next-treatment recommendations.Applications are grouped by product, content, or service recommendation.
  • Applications: Most existing SBRS work focuses on conventional applications, especially e-commerce next-item or next-basket recommendation.Other conventional domains include news, entertainment, and tourism.
  • Applications: Emerging applications in finance and healthcare remain early-stage but are described as promising for trading, portfolio, and personalized-treatment recommendations.
  • Algorithms and datasets: The survey summarizes representative open-source SBRS algorithms to facilitate empirical analysis.
  • Algorithms and datasets: For raw POI data without session structure, researchers may manually define sessions from a user’s check-ins within a single day.
  • Algorithms and datasets: It also compiles 13 publicly accessible real-world datasets spanning domains from e-commerce to POI recommendation.These datasets have varied characteristics and provide test beds for SBRS evaluation.

10 PROSPECTS AND FUTURE DIRECTIONS

The survey identifies future directions for SBRSs around integrating general preferences and context, handling cross-domain and richer behavior patterns, enforcing session constraints, and using interactive feedback. These directions reflect unresolved challenges in modeling diverse session data and user goals.

  • General Preference Learning: SBRSs need to incorporate users’ long-term general preferences alongside session-based signals, using explicit or implicit preference data.Explicit data may include ratings, while implicit data includes views, clicks, cart additions, and purchases.
  • Context-aware Session-based Recommendation: Contextual factors such as time, weather, location, season, and popularity trends can substantially affect recommendation performance.Existing work has incorporated time gaps, time of day, recent popularity, recently viewed items, and discounts.
  • Cross-domain Recommendation: Cross-domain SBRSs must either transfer knowledge from auxiliary domains or model heterogeneous dependencies among domains within shared sessions.The multi-domain setting treats items from different domains equally and allows each domain to serve as a recommendation target.
  • More Behavior Patterns: Future models should discover and leverage repeat, periodic, and other behavior patterns beyond basic co-occurrence or sequential interactions.These patterns are less frequent and harder to identify, and their influence on final choices remains unclear.
  • Constraints in Session Generation: SBRSs should generate sessions that respect item constraints, including complementarity and duplication, while jointly optimizing user goals and prediction accuracy.Different users may require different constraints, making constraint selection and joint optimization difficult.
  • Interactive Feedback: Interactive feedback from actions such as viewing, adding to cart, or skipping can support adaptive recommendation strategies during a session.These actions represent different forms of user feedback in continuous interaction with the platform.

11 CONCLUSIONS

The paper presents a systematic survey of SBRS research, organizing the field through a unified framework, problem statement, and classification scheme. It analyzes session-data characteristics and challenges, reviews representative approaches, and identifies open issues and future directions.

  • Survey Scope and Organization: The survey reviews notable SBRS research and organizes existing work into three sub-areas through a unified framework.It also proposes a unified problem statement to reduce confusion and inconsistency in the field.
  • Data Characteristics and Challenges: The paper analyzes session-data characteristics and the corresponding challenges they create for SBRSs.The review connects data properties with research challenges across the surveyed literature.
  • Approach Taxonomy: A classification scheme organizes existing SBRS approaches and highlights critical technical details for each approach class.The scheme is intended to support organization and clustering of representative methods.
  • Future Directions: The survey discusses pressing open issues and promising directions for future SBRS research.It positions the field as active, with newly developed techniques and emerging approaches continuing to appear.
Loading 1902.04864v3…