PaperlayerView exact revision on arXiv

Source-linked AI summary

Explainable Recommendation: A Survey and New Perspectives

Yongfeng Zhang, Xu Chen

arXiv:1804.11192v10cs.IRcs.AIcs.MM

TL;DR

Explainable recommendation addresses how recommender systems can provide understandable reasons for their recommendations rather than only high-quality outputs. This survey synthesizes the field’s history, explanation types, models, evaluations, applications, and broader research connections, while identifying fidelity and context-awareness as limitations or open boundaries. It concludes that several areas, including knowledge graphs, deep learning, natural language processing, user behavior analysis, and cognitive foundations, can advance the field.

  • Problem

    Recommender systems need explanations that clarify why items are recommended, but complex models can make explanation fidelity difficult to establish.

  • Method

    The survey organizes explainable recommendation through a historical review, explanation and model taxonomies, evaluation methods, applications, and connections to broader IR and AI/ML research.

  • Results

    The survey reviews user/item-based, content-based, textual, visual, and social explanations alongside MF-, topic-, graph-, deep-learning-, knowledge-, mining-, and post-hoc models.

  • Takeaways & Limitations

    Knowledge graphs, deep learning, natural language processing, user behavior analysis, model aggregation, logical reasoning, conversational systems, and cognitive foundations are identified as directions for advancing explainable recommendation.

  • Takeaways & Limitations

    Most current explainable recommendation models are static, so context-aware explanations remain an extensively unexplored direction.

Abstract

from arXiv · show

Explainable recommendation attempts to develop models that generate not only high-quality recommendations but also intuitive explanations. The explanations may either be post-hoc or directly come from an explainable model (also called interpretable or transparent model in some contexts). Explainable recommendation tries to address the problem of why: by providing explanations to users or system designers, it helps humans to understand why certain items are recommended by the algorithm, where the human can either be users or system designers. Explainable recommendation helps to improve the transparency, persuasiveness, effectiveness, trustworthiness, and satisfaction of recommendation systems. It also facilitates system designers for better system debugging. In recent years, a large number of explainable recommendation approaches -- especially model-based methods -- have been proposed and applied in real-world systems. In this survey, we provide a comprehensive review for the explainable recommendation research. We first highlight the position of explainable recommendation in recommender system research by categorizing recommendation problems into the 5W, i.e., what, when, who, where, and why. We then conduct a comprehensive survey of explainable recommendation on three perspectives: 1) We provide a chronological research timeline of explainable recommendation. 2) We provide a two-dimensional taxonomy to classify existing explainable recommendation research. 3) We summarize how explainable recommendation applies to different recommendation tasks. We also devote a chapter to discuss the explanation perspectives in broader IR and AI/ML research. We end the survey by discussing potential future directions to promote the explainable recommendation research area and beyond.

Introduction

Explainable recommendation addresses the “why” of personalized recommendation by pairing recommendations with understandable explanations for users or system designers. This survey situates the area within recommender-system research, organizes its methods, and traces its historical development and broader relevance.

  • Explainable Recommendation: Explainable recommendation supplements personalized recommendation results with explanations that clarify why items are recommended.Explanations may be generated by transparent models or added post hoc to black-box decisions.
  • Importance: Explainability matters because it can support transparency, persuasiveness, effectiveness, trustworthiness, satisfaction, and system diagnosis, while remaining connected to broader explainable AI research.The survey highlights explainable deep models as a direction relevant to both recommendation and fundamental explainable machine learning.
  • Position in Recommender Systems: The survey positions explainable recommendation as the “why” problem alongside when, where, who, and what recommendation research.These categories correspond to time-aware, location-based, social, application-aware, and explainable recommendation.
  • Scope: Explainable recommendation includes transparent models, post-hoc explanation methods, and human-centered approaches for delivering explanations and studying user interaction.The scope therefore extends beyond model construction to explanation presentation and human-computer interaction.
  • Historical Development: Explainable recommendation evolved from early content-based and collaborative-filtering systems toward explicitly explainable and model-based approaches.Earlier systems offered some intuitive explanations through item or user similarity, while later work designed algorithms to output recommendations and explanations together.
  • Method Taxonomy: The survey classifies existing methods using two orthogonal dimensions: explanation information source or display style and algorithmic mechanism.Each combination defines a sub-direction, helping readers understand relationships among explainable recommendation methods.

Information Source for Explanations

Recommendation explanations draw on varied information sources and display styles, ranging from relevant users and items to features, charts, text, and social information. These explanations support different recommendation mechanisms and can improve users’ understanding, trust, and responses to recommendations.

  • Explanation sources and styles: Explanations may use relevant users, relevant items, features, textual reviews, images, social friends, statistical charts, or reasoning rules.They can also be presented as sentences, word clouds, highlighted image regions, histograms, or pie charts.
  • Relevant-user explanations: User-based collaborative filtering explains recommendations through similar users who rated the recommended item positively.Neighbor ratings may be displayed directly, including their rating distributions and similarity to the target user.
  • Relevant-item explanations: Item-based collaborative filtering explains a recommendation by linking it to items the user previously liked.Relevant-item explanations show highly rated prior items as evidence for the recommended item.
  • User benefits: User studies identified transparency, scrutability, trustworthiness, effectiveness, persuasiveness, efficiency, and satisfaction as benefits of recommendation explanations.Providing appropriate explanations was reported to benefit recommender systems across these seven perspectives.
  • Trade-offs and social explanations: Relevant-item explanations are often more intuitive, while relevant-user explanations may be less convincing and can raise privacy concerns.Social explanations use friends’ public interests as an alternative source of explanatory information.
  • Feature and sentence explanations: Feature-based explanations clarify why item features are relevant and can improve recommendation effectiveness, trust, and satisfaction.Other approaches use templates, opinion words, or concept graphs to generate usable textual explanations, while recommendation and dis-recommendation explanations can improve persuasiveness and conversion rate.

Explainable Recommendation Models

Explainable recommendation research distinguishes explaining the recommendation method from explaining the recommendation result, using intrinsic interpretable models or separate post-hoc models. The survey then situates these approaches alongside machine-learning methods used for personalized recommendation.

  • Research philosophies: Explainable recommendation can target either the interpretability of recommendation methods or the explainability of recommendation results.Method-focused approaches design interpretable models, while result-focused approaches treat the recommendation model as a black box and explain its outputs separately.
  • Research philosophies: Intrinsic approaches use transparent decision mechanisms that naturally support explanations of model decisions.The survey identifies factorization, topic modeling, graph-based, deep-learning, knowledge-based, and rule-mining approaches as examples of this philosophy.
  • Research philosophies: Post-hoc or model-agnostic approaches generate explanations for recommendations produced by a black-box model.These approaches focus on explaining recommendation results rather than making the underlying recommendation process interpretable.
  • Machine-learning background: The survey reviews machine learning for personalized recommendation before presenting explainable recommendation models.The background includes latent-factor methods, point-wise prediction, pair-wise ranking, list-wise learning, and deep-learning approaches.
  • Machine-learning background: Latent-factor methods learn low-dimensional user and item representations to predict missing ratings, while ranking methods learn item orderings from implicit feedback.Matrix factorization is described for explicit feedback, and Bayesian Personalized Ranking is given as an example for positive versus negative items.

3.2. Factorization Models for Explainable Recommendation

Factorization-based explainable recommendation makes latent representations more interpretable by aligning factors with explicit features, modeling feature preferences, or tracing predictions to supporting data. These models use reviews, tensors, attention, structured data, and post-hoc influence analysis to produce personalized explanations.

  • Motivation: Latent factor models are difficult to interpret because the meanings of their user and item embedding dimensions are not explicit.The models assume dimensions represent factors affecting decisions, but the exact meaning of each factor is unknown.
  • Explicit feature models: Explicit Factor Models align latent factorization dimensions with product features extracted from reviews, making prediction procedures trackable.The resulting explanations connect a recommendation to a user’s interest in a feature and the product’s performance on that feature.
  • Explicit feature models: EFM can also generate dis-recommendations by identifying features important to the user on which a product performs poorly.The model was extended to represent users’ favorite features dynamically at daily resolution because feature preferences may change over time.
  • Tensor and multi-task models: Tensor-factorization extensions model user-item-feature interactions, and multi-category learning helps alleviate data sparsity.A joint tensor framework can combine preference modeling for recommendation with opinionated content modeling for explanation.
  • Attention and post-hoc models: Attention-driven factor models learn item-specific user attention distributions over features, which also serve as recommendation explanations.This addresses the limitation of assuming that each user has one global feature-preference distribution.
  • Attention and post-hoc models: Influence analysis explains latent-factor predictions by tracing them back to training data and producing neighbor-style explanations.This approach incorporates interpretability into latent-factor models through analysis of the relationship between user history and recommendations.
  • Structured extensions: Factorization-based methods can extract features and sentiments for feature recommendations, or use regression-tree structures to explain learned latent factors.These approaches extend factorization with review-derived sentiment information or structured tree guidance.
  • Relevant-user explanations: Relevant-user explanation models use only the user-item rating matrix and explain recommendations through similar users’ purchasing behavior.Explainable Matrix Factorization adds an explainability regularizer to the matrix-factorization objective.

3.3. Topic Modeling for Explainable Recommendation

This section contrasts explainable recommendation with and without external data support, highlighting rating-matrix methods that generate relevant-user explanations. Explainable Matrix Factorization explains an item through the purchasing behavior of users similar to the target user.

  • External-data support: Explainable recommendation models may use external information or rely solely on the user-item rating matrix.The survey presents both settings as alternatives for generating relevant-user or relevant-item explanations.
  • Relevant-user explanations: Explainable Matrix Factorization explains a recommended item by stating that many users similar to the target user purchased it.The model adds an explainability regularizer to the matrix-factorization objective to produce relevant-user explanations.

3.3 Topic Modeling for Explainable Recommendation

Topic-modeling approaches use review text and related information to represent user preferences, item recommendability, aspects, sentiments, and viewpoints. They generate explanations through topical words, aspect sentiment, or socially grounded opinion structures.

  • Topic-based explanations: Topic-modeling approaches commonly use textual reviews to generate intuitive explanations as topical word clouds.The survey groups these methods around explanations derived from topics learned from review content.
  • Latent topics: The HFT model links latent-factor dimensions with LDA topics to improve rating prediction and identify topics associated with user ratings.Projecting a user’s latent vector into topic space helps reveal significant topics that the user likes.
  • Latent topics: A unified topical recommendability space represents users and items, allowing explanations based on topic words with high recommendability scores.User preference and item recommendability distributions are integrated into a latent-factor framework for prediction.
  • Aspect and sentiment models: FLAME learns personalized user preferences for item aspects from reviews and explains hotel recommendations with aspect word clouds.The word-cloud size is proportional to the sentiment associated with each aspect.
  • Aspect and sentiment models: Probabilistic models can jointly represent sentiment, aspects, and regions for POI recommendation, using topical-aspect preferences as explanations.These explanations include the sentiment of each topical cluster.
  • Social explanations: Social-collaborative viewpoint regression defines viewpoints as concept-topic-sentiment tuples from reviews and trusted social relations.The viewpoints serve as explanations and are modeled probabilistically to improve recommendation prediction accuracy.

3.4 Graph-based Models for Explainable Recommendation

Graph-based explainable recommendation models represent user–item relationships as graphs and derive explanations from graph structure, collaborative patterns, or learned decision rules.

  • Tripartite and co-clustering models: TriRank models user–item–aspect ternary relations in a tripartite graph and explains recommendations using top-ranked aspects matching the user and item.The algorithm applies smoothness and fitting constraints for node ranking and personalized recommendation.
  • Tripartite and co-clustering models: Overlapping user–item co-clustering generates explanations from similar users’ purchase histories and similar items’ properties.A recommendation may receive multiple user-based and item-based explanations when a pair belongs to multiple co-clusters.
  • Rule-enhanced embeddings: Tree-enhanced embeddings combine explicit decision rules from side-information cross features with embedding-based generalization.The tree learns interpretable rules, while the embedding model incorporates those features to generalize to unseen users or items.
  • Social recommendation: Graph-based explanations also support social recommendation by using friends with similar preferences to explain recommended products.UniWalk exploits both ratings and social-network structure for explainable recommendations.

3.5 Deep Learning for Explainable Recommendation

Deep learning approaches produce explainable recommendations through attention, natural-language generation, visual highlighting, memory, and capsule-based representations across several recommendation tasks.

  • Deep learning approaches: Deep explainable recommendation spans CNNs, RNNs/LSTMs, attention mechanisms, memory networks, capsule networks, and other representation-learning techniques.These methods are applied to rating prediction, top-n recommendation, and sequential recommendation.
  • Attention and review-based explanations: Dual local and global attention selects important review words for rating prediction and highlights them as recommendation explanations.The architecture learns different attention weights over review text associated with users and items.
  • Open issue: fidelity: Deep explainable models can improve explainability and give users greater control over recommendation results.The survey identifies explanation fidelity as an open issue because complex models may not faithfully reflect the mechanisms generating recommendations.
  • Natural-language explanations: Character-level RNNs generate natural-language explanations conditioned on user ratings, while crowdsourcing selects user-tailored explanations from human-written content.Generated explanations can vary with model parameters, whereas crowd-sourced explanations are selected according to the user’s interests.
  • Attention and review-based explanations: Attention-based review selection provides high-quality reviews as explanations, helping users examine product information more quickly.The model learns review usefulness rather than generating explanations from predefined templates.
  • Other explainable deep models: Other deep models explain recommendations by highlighting visual regions, tracing influential historical items, or identifying informative item-aspect–user-viewpoint logic units.These approaches jointly model images and reviews, attend over memory slots, or use sentiment capsules for rating prediction and explanation.

3.6 Knowledge Graph-based Explainable Recommendation

Knowledge graphs provide rich user–item and entity information that supports intuitive, tailored explanations alongside recommendations.

  • Knowledge-graph foundations: Knowledge graphs contain rich information about users and items, making them useful for generating intuitive and more tailored recommendation explanations.Recent approaches explore knowledge graphs specifically for explainable recommendation.
  • Knowledge-graph foundations: Personalized PageRank jointly ranks movies and knowledge-graph entities to produce recommendations together with ranked explanatory entities.The method is described in a movie recommendation setting.
  • Knowledge-graph foundations: Knowledge-graph embeddings represent users, items, entities, and relations, while shortest paths connect users to recommended items as explanations.The graph includes relations such as purchases and category membership.

3.6. Knowledge Graph-based Explainable Recommendation

Knowledge-graph recommendation methods propagate preferences, model sequential and attribute-based interests, induce relational rules, and use path reasoning to generate explainable recommendations.

  • Graph reasoning and paths: Knowledge-graph embeddings recommend items through purchase relations and explain them with paths connecting users to recommended items.Paths may pass through shared entities such as mentioned words or brands.
  • Preference propagation: Ripple Network iteratively propagates a user’s potential interests through knowledge-graph links from historically clicked items.Multiple activated preference ripples combine into a user preference distribution.
  • Sequential recommendation: A recurrent neural network captures sequential item preferences while a key-value memory network captures attribute-based preferences from item knowledge.The combined model also predicts the importance of different values for critical attributes.
  • Graph reasoning and paths: Policy-guided path reasoning learns to navigate a knowledge-graph environment from users toward potential items, using the resulting path as the explanation.The approach addresses the computational cost of enumerating all user–item paths in large knowledge graphs.
  • Sequential recommendation: Sequential knowledge-graph models capture dynamic user preferences by modeling interactions over multimodal or ordinary knowledge graphs.These methods extend beyond conventional item-level sequential modeling.
  • Rule-guided recommendation: Joint rule induction and neural recommendation combine mined multi-hop relational patterns with a rule-guided neural model.The two modules are designed to complement each other in generating explainable recommendations.

3.7. Rule Mining for Explainable Recommendation

Rule-mining methods generate explainable recommendations by extracting associations among users, items, or behaviors, often yielding direct, self-explanatory reasons.

  • Rule mining is useful for explainable recommendation because the resulting rules can directly explain why items are recommended.Association rule mining is the most frequently used rule-mining technique in this context.
  • Association rule mining supports practical recommendation systems, including web-page, web-shop, conversational, and task-based recommendation.
  • Personalized association rule mining extracts rules for a target user and uses user-item associations to produce recommendations.
  • YouTube used co-watched video pairs within user sessions to calculate relatedness scores and personalize recommendations from each user’s seed videos.

3.8. Model Agnostic and Post Hoc Explainable Recommendation

Model-agnostic and post-hoc methods explain black-box recommendations with a separate explanation model, trading exact fidelity for broad applicability across recommendation models.

  • Post-hoc approaches generate explanations after recommendations using an explanation model separate from the black-box recommendation mechanism.The explanations are decoupled from the recommendation model but should still report true statistical information.
  • Post-hoc explanations can treat arbitrary recommendation models as black boxes and derive association rules from users’ model-input histories and outputs.
  • Singh and Anand trained an explainable tree-based model on ranking labels produced by a black-box ranker to explain ranking lists.Their method explains rankings pairwise, whereas Peake and Wang’s approach explains recommendations pointwise.
  • Reinforcement learning can generate sentence explanations by rewarding agents for predicting black-box ratings and producing explanations with desirable presentation quality.The agents optimize expected rewards while interacting with the recommendation model as their environment.
  • FIA uses influence analysis to trace latent-factor model predictions back to training data for post-hoc explanation.
  • Post-hoc methods are flexible across recommendation models, but their explanations may not precisely follow the mechanism that produced each recommendation.

3.9 Summary

The survey organizes explainable recommendation around transparent and model-agnostic approaches, reflecting whether explanations arise from the recommendation method or are generated for its results.

  • The survey covers matrix and tensor factorization, topic modeling, graph-based, deep learning, knowledge-based, rule-mining, and post-hoc approaches.
  • Explainability can concern either the recommendation method itself or the recommendation results, corresponding to model-intrinsic and model-agnostic approaches.
  • These two modeling philosophies are connected to contrasting human decision patterns: reasoning first and explaining naturally, or deciding first and seeking an explanation afterward.

Evaluation of Explainable Recommendation

Explainable recommendation evaluation combines conventional recommendation metrics with offline, online, user-study, and case-study assessments of explanation quality and user impact.

  • Evaluation seeks recommendation performance comparable to conventional methods while also assessing explainability.
  • Rating prediction can use MAE or RMSE, while top-n recommendation can use precision, recall, F-measure, and NDCG.
  • Offline explanation evaluation is easier to implement than online evaluation because online studies require accessible real-world data and users.Online evaluation is encouraged but not always required.
  • User studies measure explanation effects through participant responses about recommendation understanding, effectiveness, mood compatibility, trust, or willingness to choose items.
  • Studies found that tag preference and relevance affect justification, effectiveness, and mood compatibility, while explanations can enhance consumers’ initial trusting beliefs.
  • Online experiments evaluate persuasiveness, effectiveness, efficiency, and satisfaction, often using click-through, cart-addition, or agreement measures.
  • Case studies identify one-to-multiple and one-to-one purchase behaviors as patterns that explain sequential recommendations.These patterns clarify why an item is recommended and how it matches prior purchases.
  • Explainable recommendations can support efficient and informed e-commerce decisions, and appropriate explanations can increase acceptance of location recommendations and click-through rates for generated song reasons.

Open Directions and New Perspectives

The survey identifies open directions spanning explainable deep learning, domain knowledge, heterogeneous information, context, explanation diversity, natural language, and fairness. It also frames explainability as a broader trade-off between transparent reasoning and the predictive strengths of complex models.

  • Methodology: Explainable deep learning remains at an early stage, with current systems generating explanations from attention over text, images, or video frames.The explainability of the deep model itself remains unresolved because hidden layers often lack intuitive meanings.
  • Methodology: Knowledge graphs and graph embeddings can support domain-specific explanations through knowledge reasoning and conversational recommendation.Such systems can connect recommended items to explicit entities and relations, such as actors and movies.
  • Methodology: Heterogeneous information sources remain underused for explainability despite their widespread use for improving retrieval and recommendation performance.Open tasks include multimodal explanations, transfer learning across sources, and cross-domain explanation.
  • Evaluation: Most explainable recommendation models are static, although changing preferences and context suggest that explanations should adapt to time and location.Different models also produce explanations suited to different purposes, while diversified explanations can benefit user satisfaction.
  • Methodology: Natural-language explanations remain an early research direction because review content contains noise and explanations must be personalized and informed by multimodal or external knowledge.The survey points toward adapting pretrained language models and integrating images, knowledge graphs, and sentiment.
  • Broader impacts: Explainability and fairness form an important open problem, while the field also faces a transparency trade-off between interpretable symbolic methods and more robust, effective complex models.Collaborative filtering followed a similar path from transparent user/item methods toward latent models whose outputs arise from implicit similarity matching rather than explicit reasoning.

Conclusions

The survey traces explainable recommendation from transparent early methods to complex models that improve recommendation performance but reduce transparency. It consolidates explanation types, model families, evaluation methods, applications, and future perspectives, positioning explainability as important for recommendation and broader AI.

  • Conclusions: Complex latent-factor and deep representation models improved recommendation performance but made recommendation outputs and mechanisms harder for users and designers to understand.This creates separate explainability needs for recommendation results and the algorithms that produce them.
  • Conclusions: Explainability is especially important in decision-making settings such as healthcare, where users need to understand provided results and how to act on them.The survey gives medical diagnosis as an example requiring comprehensive and explainable recommendations or retrieval.
  • Conclusions: The survey consolidates a history of explainable recommendation, explanation types, model families, evaluation methods, and applications across e-commerce, POI, social, and multimedia recommendation.Its outlook includes knowledge graphs, deep learning, natural language processing, user behavior analysis, model aggregation, logical reasoning, conversational systems, and cognitive foundations.
  • Conclusions: Explainable recommendation is part of a broader Explainable AI agenda spanning deep learning, computer vision, autonomous driving, and natural language processing.The survey emphasizes the relevance of explainability research to information retrieval and recommendation systems.
Loading 1804.11192v10…