Source-linked AI summary
Continual Recommender Systems
Hyunsik Yoo, SeongKu Kang, Hanghang Tong
TL;DR
Continual recommender systems must adapt to changing users and items while retaining prior knowledge, yet existing tutorials do not address recommendation-specific challenges. This tutorial reviews background, methods, and practical settings, and identifies future directions including trustworthiness and foundational models.
Problem
Method
The tutorial systematically reviews continual recommendation, covering background, replay- and regularization-based approaches, and practical resource-constrained and sequential settings.
Source: §1 · p. 1, §4.1 · p. 2, §4.2 · p. 2Results
The tutorial organizes prior work into experience-replay and regularization strategies and surveys practical environments and future research directions.
Source: §1 · p. 1Takeaways & Limitations
The tutorial offers researchers and practitioners an overview of prior efforts and a walk-through of adapting to evolving preferences and new users or items.
Source: §2 · p. 2Two-sided fairness remains relatively underexplored, and fairness maintenance is challenging amid evolving behavior and accuracy trade-offs.
Source: §4.5 · p. 3
Abstract
from arXiv · showhide
Modern recommender systems operate in uniquely dynamic settings: user interests, item pools, and popularity trends shift continuously, and models must adapt in real time without forgetting past preferences. While existing tutorials on continual or lifelong learning cover broad machine learning domains (e.g., vision and graphs), they do not address recommendation-specific demands-such as balancing stability and plasticity per user, handling cold-start items, and optimizing recommendation metrics under streaming feedback. This tutorial aims to make a timely contribution by filling that gap. We begin by reviewing the background and problem settings, followed by a comprehensive overview of existing approaches. We then highlight recent efforts to apply continual learning to practical deployment environments, such as resource-constrained systems and sequential interaction settings. Finally, we discuss open challenges and future research directions. We expect this tutorial to benefit researchers and practitioners in recommender systems, data mining, AI, and information retrieval across academia and industry.
1 Introduction
Recommender systems face continuously changing users and item pools, making full-dataset retraining too slow for rapid updates. Continual learning must balance adaptation with preserving prior knowledge, and this tutorial reviews methods and future directions for recommendation settings.
1 Introduction
- Changing user preferences, behaviors, and available items can quickly make recommender models stale, while full-dataset retraining is often too slow and resource-intensive for rapid updates.Fine-tuning on new data is described as a more efficient alternative to full retraining.Source: §1 · p. 1
- Continual learning must balance plasticity with stability; naïve fine-tuning can overwrite prior knowledge and degrade recommendations, especially for returning users and long-tail items.The balance may need personalization across users, while deployment can also be constrained by memory, computation, or energy budgets.Source: §1 · p. 1
- The tutorial systematically reviews continual recommender-system research, covering background, replay- and regularization-based methods, practical environments, and open directions.Its practical settings include resource-constrained and sequential-interaction environments.Source: §1 · p. 1
2 Target Audience and Benefits
The tutorial targets researchers and practitioners across data mining, AI, information retrieval, and related fields. It connects continual learning to recommendation's changing data, user preferences, and new users and items.
2 Target Audience and Benefits
- The intended audience includes researchers and practitioners in data mining, artificial intelligence, information retrieval, and related fields.The stated prerequisites are basic probability, linear algebra, and machine learning, with no prior familiarity with continual-learning algorithms required.Source: §2 · p. 1
- Recommendation quality matters for user satisfaction and engagement, and can affect business revenue across online platforms.The tutorial notes that personalization techniques may also apply to user modeling in information retrieval, graph mining, and natural language processing.Source: §2 · p. 2
- Continual learning offers an overview of fine-tuning on arriving data to adapt to changing preferences and new users or items when frequent retraining is infeasible.The tutorial provides a friendly walk-through of prior efforts for audience understanding.Source: §2 · p. 2
3 Outline and Timeline
The half-day tutorial is organized into five parts, moving from background through methods and practical settings to future directions. Its agenda allocates 3 hours plus breaks.
3 Outline and Timeline
- The tutorial lasts 3 hours plus breaks and allocates 30 minutes to introduction and background.The opening part covers problem definitions and settings, key challenges, and applications.Source: §3 · p. 2
- Experience-replay and regularization-based methods each receive 35 minutes, covering sample selection and replay enhancement, then knowledge and temporal regularization and personalization.A 5-minute short break follows these parts.Source: §3 · p. 2
- Beyond-traditional settings and future directions each receive 35 minutes, followed by a 5-minute conclusion.The agenda includes resource-constrained and sequential interaction environments, trustworthiness, foundational models, and unified recommendation and search.Source: §3 · p. 2
4 Tutorial Content
The tutorial moves from continual-recommendation challenges to replay and regularization strategies, then examines practical settings and open research directions. It highlights personalization, resource constraints, trustworthiness, foundational models, and unified search and recommendation.
4.1 Part I: Introduction and Background
- The introduction and background cover training and inference settings, forgetting, changing preferences and item popularity, new users and items, personalized stability–plasticity, and practical use cases.These topics frame the tutorial's overview of continual recommendation.Source: §4.1 · p. 2
4.2 Part II: Experience-Replay-based Methods
- Experience replay reuses past interactions, with methods addressing which samples to retain and how to use them during training or prediction.Selection approaches include frequency, past errors, model-training influence, and encoded user-item-context features; enhancement approaches include specialized losses, stored prior-model outputs, and adaptation modules.Source: §4.2 · p. 2
4.3 Part III: Regularization-based Methods
- Regularization constrains current parameters using knowledge from different time steps; methods vary in the knowledge regularized, its temporal direction, and whether weights are personalized.Approaches regularize user representations using local or global structure and interaction-network self-information, or maximize mutual information between past and current GNN parameters; forward knowledge can promote plasticity, while user dynamics inform personalized weights.Source: §4.3 · pp. 2–3
4.4 Part IV: Beyond Traditional Settings
- Resource-constrained and sequential-interaction deployments pose additional adaptation challenges beyond traditional settings.The tutorial covers selective replay and lightweight-model adaptation via distillation under resource limits, while noting that dedicated continual approaches for sequential-pattern models remain underexplored.Source: §4.4 · p. 3
4.5 Part V: Future Directions
- Future directions include trustworthiness, continual adaptation of foundational models, and unified models for recommendation and search.The tutorial identifies adversarial robustness, diversity, and explainability as further trustworthiness concerns.Source: §4.5 · p. 3
- Two-sided fairness remains relatively underexplored, and maintaining fairness is practically challenging amid evolving behavior and trade-offs with accuracy.For foundational models, cited continual-LoRA efforts either focus on other domains or overlook personalization of the stability–plasticity trade-off.Source: §4.5 · p. 3
5 Related Tutorials
The tutorial differs from related tutorials by centering recommendation-specific continual learning, including real-time updates and personalization. It distinguishes this focus from general continual-learning and graph-learning frameworks and from cross-domain recommendation.
5 Related Tutorials
- General continual-learning tutorials cover broad methods and benchmarks but do not target recommendation-specific challenges such as evolving interactions, personalization, or user-preference adaptation.The cited tutorials focus on general machine-learning domains and theory.Source: §5 · p. 3
- The cross-domain recommendation tutorial presents the 3C principle for integrating knowledge across domains, whereas this tutorial focuses on real-time, single-domain updates and personalization.The 3C principle means Complement, Composite, Context.Source: §5 · pp. 3–4
- Continual graph-learning tutorials address dynamic-graph frameworks and techniques such as replay and regularization, while this tutorial targets recommendation goals including preference shifts, popularity changes, and personalized stability–plasticity trade-offs.The graph-learning tutorials also discuss dynamic benchmarks.Source: §5 · pp. 3–4
6 Presenter Biography
The tutorial is presented by Hyunsik Yoo, SeongKu Kang, and Hanghang Tong, whose biographies describe their positions and research backgrounds. The manuscript also discloses the use of GPT-4 for grammar and clarity, with suggestions reviewed and approved by the authors.
6 Presenter Biography
- Hyunsik Yoo is a fourth-year Ph.D. student at the University of Illinois Urbana-Champaign whose research includes recommender systems and graph mining.His work has appeared at KDD, SIGIR, TheWebConf, WSDM, and ICML.Source: §6 · p. 4
- SeongKu Kang is an Assistant Professor at Korea University whose research interests include data mining, recommender systems, and information retrieval.He previously worked as a postdoctoral researcher at the University of Illinois Urbana-Champaign.Source: §6 · p. 4
- Hanghang Tong is a professor at the University of Illinois Urbana-Champaign whose research focuses on large-scale data mining for graphs and multimedia.His biography reports more than 300 refereed articles and several awards.Source: §6 · p. 4
- OpenAI GPT-4 was used for manuscript grammar and clarity, and the authors manually reviewed and approved its suggestions.Source: §6 · p. 4