PaperlayerView exact revision on arXiv

Source-linked AI summary

Recent Developments in Recommender Systems: A Survey

Yang Li, Kangbo Liu, Ranjan Satapathy, Suhang Wang, Erik Cambria

arXiv:2306.12680v1cs.IRcs.AIcs.LG

TL;DR

Recommender systems are surveyed across personalized and group approaches, with attention to current challenges and evaluation. The paper synthesizes taxonomies, techniques, and metrics, and concludes with a broad overview of the field, including knowledge-based systems and their applications.

  • Problem

    The survey addresses the need for an overview of recommender-system approaches amid challenges involving robustness, data bias, and fairness.

    Source: §IV · p. 10

  • Method

    The paper surveys personalized and group recommender systems, knowledge-based systems, common challenges, and performance metrics.

    Source: §I · p. 2, §VI · pp. 15–16

  • Results

    The survey synthesizes recommender-system categories, techniques, challenges, and metrics into an overview of the field.

    Source: §VI · pp. 15–16

  • Takeaways & Limitations

    Knowledge-based recommender systems have applications in ecommerce, healthcare, academia, and software development.

    Source: C. Knowledge-based RS · p. 6

Abstract

from arXiv · show

In this technical survey, we comprehensively summarize the latest advancements in the field of recommender systems. The objective of this study is to provide an overview of the current state-of-the-art in the field and highlight the latest trends in the development of recommender systems. The study starts with a comprehensive summary of the main taxonomy of recommender systems, including personalized and group recommender systems, and then delves into the category of knowledge-based recommender systems. In addition, the survey analyzes the robustness, data bias, and fairness issues in recommender systems, summarizing the evaluation metrics used to assess the performance of these systems. Finally, the study provides insights into the latest trends in the development of recommender systems and highlights the new directions for future research in the field.

I. INTRODUCTION

This survey presents a broad overview of recommender systems, their taxonomies, implementation, and evaluation, extending beyond prior surveys focused on particular domains. It covers personalized and group recommendation, alongside robustness, data bias, fairness, and evaluation metrics.

II. PERSONALIZED RECOMMENDER SYSTEMS

  • Personalized systems predict an individual’s next item from past preferences and behavior, while group systems combine users’ collective preferences to alleviate conflicts.Personalized systems are categorized as collaborative filtering, content-based, and hybrid; the survey also examines knowledge-based systems.Source: §I · p. 1, §II · p. 2

A. Collaborative Filtering Based RS

  • Collaborative filtering uses other users’ behavior and can be divided into memory-based and model-based approaches.Model-based approaches employ mathematical models, while memory-based approaches use user profiles more directly.Source: §I · p. 1, A. Collaborative Filtering Based RS · p. 2

I. INTRODUCTION

  • Group recommendation is described as more difficult because it must account for multiple users’ interests, and the survey focuses on popular memory-based and model-based categories.It excludes some other types, including multicriteria, risk-aware, and mobile recommender systems.Source: §I · p. 1
  • The survey classifies evaluation metrics as rating-based indicators, which assess predicted ratings, or item-based indicators, which assess predicted item lists.This taxonomy is presented as covering most existing evaluation indicators.Source: §I · pp. 1–2
  • Earlier surveys covered particular domains, including contextual-aware recommendation, deep learning, knowledge graphs, and graph neural networks.The survey distinguishes its broader coverage from these focused overviews.Source: §I · p. 2
  • The survey offers a broad overview of recommender-system problems and techniques, and introduces a taxonomy alongside evaluation metrics and challenges.It addresses robustness, data bias, and fairness, and is positioned as broader than surveys focused on specific domains.Source: §I · p. 2

1) Memory-based CF:

Collaborative filtering ranges from memory-based nearest-neighbor methods to model-based neural and graph approaches. Memory-based methods are simple but face sparsity and scaling challenges, while graph-neural models show strong MovieLens results in the surveyed comparison.

1) Memory-based CF:

  • Memory-based CF computes user or item similarity with nearest neighbors, then recommends according to the sorted similarity values.It is described as simple and efficient, using historical behavior to capture preferences.Source: 1) Memory-based CF · p. 2
  • Sparse interaction matrices make memory-based predictions more error-prone, while growing user or item spaces make similarity calculations costly and difficult to scale.The survey lists PCC, cosine similarity, MSD, PIP, Jaccard similarity, and PSS as possible similarity measures.Source: 1) Memory-based CF · p. 2

2) Model-based CF:

  • Neural-based CF maps user and item representations, and review text when available, into a continuous vector space to predict ratings.The model parameters are optimized to minimize prediction error, and the predicted rating may be refined with a factorization-based method.Source: 2) Model-based CF · p. 3
  • Factorization machines model user-item interactions with feature vectors and are presented as efficient for sparse data, with complexity O(dk) rather than O(d^2).The rank hyper-parameter k can be adjusted according to the data type.Source: 2) Model-based CF · p. 3
  • Model-based CF also incorporates review features through CNNs, attention, aspect representations, and capsule networks, with some surveyed work targeting explanations.Examples include DeepCoNN, aspect-based neural recommender ANR, and capsule-network review modeling.Source: 2) Model-based CF · p. 3
  • On MovieLens, the two combined graph-neural methods, achieve the best reported NDCG@K and Recall@K results among the seven methods.The survey says neighborhood-enhanced contrastive learning slightly outperforms the combined incremental graph-convolution methods; basic graph collaborative-filtering models are comparable, with slightly ahead of and [33].Source: Table I · p. 4, 2) Model-based CF · p. 4

3) Context-aware Based CF:

Context-aware CF incorporates contextual signals through filtering or direct modeling, although those signals can be complex and dynamic. Content-based systems instead use item features alongside user preferences and interaction history, including content that does not require ratings.

3) Context-aware Based CF:

  • Context may be gathered from devices or users, and recent methods can infer it from system activity rather than relying only on self-reports.Examples include geolocation and companion data from smart devices.Source: 3) Context-aware Based CF · p. 4
  • Context features can be complex and dynamic, from changing weather variables to smart-device data with as many as 600 variables.Latent context embeddings and hierarchical representations are among the surveyed approaches for managing this complexity.Source: 3) Context-aware Based CF · p. 4
  • Context enters recommendation through pre-filtering, post-filtering, or contextual modeling, with modeling integrating context directly into the recommendation algorithm.The survey notes comparatively little work on effective context use in deep-learning frameworks and describes CNN, item-splitting, autoencoder, and graph-convolution approaches.Source: 3) Context-aware Based CF · p. 5

B. Content Based RS

  • Content-based RS treats recommendation as user-specific classification using item features, user preferences, and interaction history.It can recommend from item content without ratings, using metadata or user-generated content and, increasingly, visual and multimedia features.Source: B. Content Based RS · p. 5
  • Content-based systems also connect heterogeneous side information through networks such as social graphs, bibliographic networks, and knowledge graphs.The survey describes HINs as linking multiple item types and relationships to integrate richer semantics.Source: B. Content Based RS · p. 6
  • Content-based recommendation can use clustering or ranking, with similarity measures including Kendall-Tau, Spearman, and Pearson.Source: B. Content Based RS · p. 6

C. Knowledge-based RS

Knowledge-based recommenders use domain knowledge rather than user-item interaction history, supporting recommendations in complex domains and helping address cold-start and grey-sheep challenges. The survey reviews knowledge-graph representations and applications across domains.

C. Knowledge-based RS

  • Knowledge-based RS uses domain knowledge rather than interaction history, a fit for complex purchase domains and a way to address cold-start and grey-sheep challenges.Knowledge may describe users, items, or their interactions, and can be located on either the user or item side.Source: C. Knowledge-based RS · p. 6
  • The surveyed applications span e-commerce, healthcare, academia, and software development, illustrating the range of domains where knowledge-based recommendations are used.Examples include medicine recommendation, news, travel planning, and software-development tools.Source: C. Knowledge-based RS · p. 6

1) Knowledge Graph-based RS:

  • A knowledge graph represents entities and their relationships as triples (h, r, t), and recommender systems use it through embedding, path-pattern learning, or hybrid methods.Knowledge graph embeddings can be aggregated as item-side information in latent vectors.Source: 1) Knowledge Graph-based RS · p. 7
  • TransE, TransH, TransR, TransD, and DistMult provide surveyed ways to score or map knowledge-graph entities and relations.For TransE, a low score for a triple indicates that its relation holds.Source: 1) Knowledge Graph-based RS · p. 7
  • A margin-based ranking loss contrasts correct knowledge triples with crafted incorrect triples, using γ as the margin between them.Source: 1) Knowledge Graph-based RS · p. 7
  • The survey describes external knowledge additions to KGE, including structural, visual, and textual knowledge, and methods for sparsity, noise, long-tail effects, and recall bias.Examples include TransR-based POI/entity information sharing and Path Language Modeling Recommendation.Source: 1) Knowledge Graph-based RS · p. 7

2) Path-based RS:

Path-based recommendation uses connectivity patterns in knowledge graphs to enrich user or item side information. Surveyed approaches compute semantic similarity over meta-paths and include recurrent, attention-based, and reinforcement-learning methods.

2) Path-based RS:

  • Path-based recommendation evaluates semantic similarity between users or items over knowledge-graph meta-paths, including user-user, item-item, and user-item similarity.Source: 2) Path-based RS · p. 8
  • Other knowledge-graph methods integrate association rules into matrix factorization, combine TransE embeddings with reinforcement learning, or blend path-based and propagation-based approaches –.Source: 2) Path-based RS · p. 8
  • Surveyed path-based models encode paths with recurrent networks or LSTMs, use meta-path attention, or learn recommendation paths through clustering and reinforcement learning –.The approaches include RKGE, KPRN, a guided session model, GPR, and RMS/HRec.Source: 2) Path-based RS · p. 8
  • The survey characterizes path-based methods as widely applied for using knowledge-graph connectivity, with PathSim, Hetesim, and LSTM among the similarity methods.Source: 2) Path-based RS · p. 8

D. Hybrid RS

Hybrid recommender systems combine information sources to support real-time, accurate recommendations, while group recommender systems extend recommendation to users with shared preferences. The survey describes hybrid works and a group-recommendation framework, noting that the field remains at an early stage.

D. Hybrid RS

  • Hybrid recommender systems integrate multiple data sources, including user-item interactions and contextual features, to address practical demands for real-time, accurate recommendations.The survey notes deep learning's role in fusing sources such as context and content.Source: D. Hybrid RS · p. 8
  • Most hybrid recommender works use a collaborative-filtering-based optimization model, alongside content and contextual information such as behavior, social connections, and personal characteristics.Table IV summarizes hybrid recommender systems based on model and data source.Source: D. Hybrid RS · p. 8, Table IV · p. 8

III. GROUP RECOMMENDER SYSTEMS

  • Group recommendation serves users with similar preferences, unlike the predominant individual-recommendation focus in the field.The survey describes the area as early in development, with further research needed to understand its potential and limitations.Source: §III · pp. 8–9
  • A group recommendation framework includes group creation and recommendation generation, with membership shaped by shared interests, joint activities, or a shared environment.Group creation can also be influenced by recommendation aggregation.Source: §III · p. 8

A. Memory-based Approach

The survey reviews group recommendation through preference aggregation, model-based methods, and bundled recommendations, then outlines recommender-system challenges in robustness, data bias, and fairness.

A. Memory-based Approach

  • Group preference aggregation includes result aggregation of individual recommendations and profile aggregation for a virtual user; strategies include majority, consensus, and borderline approaches.Averages can work well for similar group preferences but may yield worse recommendations when preferences are diverse.Source: A. Memory-based Approach · p. 9

B. Model-based Approach

  • Model-based group recommendation uses machine learning to represent group preferences, but performance remains unsatisfactory because group preferences are sparse and diverse.The survey describes deep learning methods that integrate external features, including social features, group descriptions, and commonsense knowledge.Source: B. Model-based Approach · pp. 9–10

C. Bundle Recommendation

  • Bundled recommendation treats two or more items as a whole, with the potential to meet user needs while reducing users' time and increasing sales revenue.The survey relates bundled recommendations to group recommendations by framing bundles as groups of items.Source: C. Bundle Recommendation · p. 10

IV. PROBLEMS IN RECOMMENDER SYSTEM

  • The survey identifies robustness, data bias, and fairness as major recommender-system challenges and reviews methods addressing them, including their strengths and limitations.It also aims to provide future research directions.Source: §IV · p. 10

A. Robustness

The survey reviews recommender-system robustness against attacks and defenses, then describes data biases arising from observed user behavior and approaches used to mitigate them.

A. Robustness

  • Adversarial attacks on recommender systems include white-box and black-box attacks, while data poisoning injects fake users and ratings to change recommendation lists.The survey also distinguishes targeted attacks, which seek a specific wrong class, from non-targeted attacks, which seek any class other than the ground truth.Source: A. Robustness · p. 11
  • Adversarial training and model distillation are commonly used defenses; adversarial training pits perturbation-seeking attackers against model parameters optimized to reduce risk.Distillation transfers knowledge from a teacher to a student model using soft labels.Source: A. Robustness · p. 11

B. Data Bias

  • Selection bias arises because users do not rate every item, producing missing-not-at-random data; observed ratings differ from the distribution of all ratings.Ranking considers all items, including those users did not rate.Source: B. Data Bias · p. 12
  • The survey discusses popularity, selection, exposure, and position biases, with debiasing techniques, score-calibration losses, and propensity scores among the approaches used to address bias.Its examples include propensity estimation and position-bias-aware click-through-rate prediction.Source: B. Data Bias · pp. 11–12

C. Fairness

The survey distinguishes user-based fairness from item-based fairness and reviews approaches to user fairness, including meta-learning, adversarial training, differential privacy, and multi-armed bandits.

C. Fairness

  • Recommender-system fairness includes protecting users from discrimination based on sensitive attributes and giving each item an equal chance to be recommended.Cold-start and long-tail problems are identified as item-based fairness cases; the long tail is also related to exposure bias.Source: C. Fairness · p. 12

1) User-Based Fairness:

  • User-based fairness approaches include meta-learning, adversarial training, federated learning, reinforcement learning, and Transformers.The cited work on CLOVER reports improving individual, group, and counterfactual fairness without degrading cold-start performance.Source: 1) User-Based Fairness · pp. 12–13

2) Item-Based Fairness:

The section surveys approaches to item-based fairness and introduces evaluation measures for recommender systems. It distinguishes rating-based metrics and describes rating prediction errors and their limitations.

2) Item-Based Fairness:

  • Item-based fairness spans causal inference, adversarial training, meta-learning, and reinforcement learning; adversarial regularization may worsen popularity deviation under unbalanced data.The survey also describes approaches aimed at reducing popularity bias, improving cold-start performance, or recommending sparse items.Source: 2) Item-Based Fairness · p. 13

V. MEASUREMENTS

  • Evaluation indicators divide into rating-based measures for rating-score outputs and item-based measures for outputs that are sets or lists of items.Rating-based measures assess closeness to true ratings; item-based measures assess whether predicted items match users’ desired items.Source: §V · p. 14

A. Rating Based Indicator

  • RMSE and MAE measure rating prediction errors, with lower values preferred, but their error terms make them sensitive to outliers.RMSE is identified as a popular measure for explicit ratings; MAE is another popular measure.Source: A. Rating Based Indicator · p. 14

B. Item Based Indicator

For item-list recommendations, the survey describes confusion-matrix measures and ranking metrics. It contrasts top-k and macro ranking measures with NDCG and NDPM’s distinct handling of relevance and ties.

B. Item Based Indicator

  • When recommendations have no ranking information, a confusion matrix can support item-based evaluation, including Precision, Recall, and F-Measure.Its outcomes distinguish recommended used items (TP), unrecommended unused items (TN), and the corresponding false-positive and false-negative cases.Source: B. Item Based Indicator · p. 14, Table V · p. 14

1) Ranking Indicators:

  • Precision@k and Recall@k evaluate only the first k items, while MAP@k and MAR@k average performance across users but require thresholding fine-grained ratings.That thresholding can cause bias and loss of fine-grained information.Source: 1) Ranking Indicators · p. 14
  • NDCG ranks highly relevant items near the top while retaining fine-grained ratings, but it cannot handle incomplete ratings without mean or median imputation.Its normalization uses ideal DCG.Source: 1) Ranking Indicators · p. 14
  • NDPM evaluates predictive rankings, including tied pairs, but does not suit systems intended to provide accurate predictive ratings.It is appropriate when users’ relative ranking information may tie two items.Source: 1) Ranking Indicators · pp. 14–15

2) Similarity:

The survey lists similarity measures used in recommender systems and discusses Pearson correlation and cosine similarity in memory-based collaborative filtering. It notes a trade-off between co-rated-item count, accuracy, and computational complexity for Pearson correlation.

2) Similarity:

  • Similarity measures include Pearson correlation, cosine similarity, mean squared difference, Jaccard, and other methods used in recent work.The survey lists these methods in its table of memory-based collaborative-filtering similarities.Source: 2) Similarity · p. 15
  • Pearson correlation ranges from −1 to 1, and its accuracy and computational complexity typically rise with the number of co-rated items.A value of 0 indicates no correlation; 1 and −1 indicate positive and negative correlation, respectively.Source: 2) Similarity · p. 15
  • Unlike Pearson correlation, cosine similarity uses all ratings and is described as possessing eigenvalue shrinkage for large eigenvalues.Missing ratings are treated as zero in the cosine-similarity formulation discussed.Source: 2) Similarity · p. 15

VI. CONCLUSION

The conclusion recaps a broad survey of recommender-system categories, challenges, and evaluation metrics. It also points to research on sentiment analysis and conversational recommendation as areas of emerging interest.

VI. CONCLUSION

  • Table VI presents similarity methods used in memory-based collaborative filtering and defines notation for ratings, co-rated items, and related quantities.The caption also identifies the median rating and average user rating notation.Source: Table VI · p. 15, §VI · p. 15
  • A cited bibliometric analysis reports that soft-computing research shifted from recommender systems toward sentiment analysis over the past two decades.The passage suggests future work such as fuzzy logic for aspect-based sentiment analysis or sentiment knowledge-base reasoning.Source: §VI · p. 15
  • The survey covers personalized and group-based recommenders, including knowledge-based systems, memory- and model-based group methods, item-oriented group recommendation, and bundled recommendation.It also reviews robustness, data bias, fairness, and rating-, ranking-, and similarity-based performance metrics.Source: §VI · pp. 15–16
  • The survey frames recommender systems as a dynamic research area shaped by diverse applications and growing data and computing resources, while noting continuing robustness, bias, and fairness challenges.Its opening conclusion describes the survey as an overview of recent developments.Source: §VI · pp. 15–16
  • The conclusion describes conversational recommendation as gaining popularity and cites ChatGPT-based and M6-based recommendation systems as examples.It says GPT technology is expected to catalyze advances in recommender systems.Source: §VI · p. 16

Related papers

Loading 2306.12680v1…