Explainable Recommendation: A Survey and New Perspectives

1.1 Explainable Recommendation

• Explainable recommendation supplies recommendation results and explanations, improving transparency, persuasiveness, effectiveness, trustworthiness, satisfaction, and system debugging.It addresses the “why” problem for both users and system designers.Source: §1.1 · p. 6

1.2 A Historical Overview

• Explainable recommendation emerged from earlier content-based and collaborative-filtering explanations, while latent-factor models improved prediction but made recommendation rationales difficult to interpret.Early examples include familiar-item explanations, user surveys, transparency studies, and industry practices such as “people also viewed” (Schafer et al., 1999; Herlocker et al., 2000; Sinha and Swearingen, 2002; Tintarev and Masthoff, 2007a). Zhang et al. (2014a) later aligned latent dimensions with explicit features in the Explicit Factor Model.Source: §1.2 · pp. 7–9

1.3 Classification of the Methods

• The survey classifies explainable recommendation methods along model type and explanation information or display style, covering approaches such as matrix factorization, topic modeling, graph models, deep learning, and knowledge reasoning.The taxonomy combines human-computer interaction and machine-learning perspectives; Table 1.1 provides an incomplete representative enumeration.Source: §1.3 · pp. 10–11, Table 1.1

1.4 Explainability and Effectiveness

• Explainability and effectiveness may require a trade-off between simple interpretable models and complex accurate models, although recent work suggests deep representation learning can support both goals.Developing explainable deep models is also presented as a direction relevant to broader explainable machine learning.Source: §1.4 · p. 13

1.5 Explainability and Interpretability

• Interpretability is one route to explainability, but explainability can also come from complex techniques such as attention mechanisms, natural-language explanations, and post-hoc explanation models.Interpretable models expose locally or globally transparent decision mechanisms, whereas non-interpretable models may be explained after or through their decisions.Source: §1.5 · p. 13, §1.6 · p. 14

1.6 How to Read the Survey

• The survey proceeds from user-interaction information sources and display styles to machine-learning models, evaluation protocols, real-world applications, and open problems and future directions.Sections 2–6 organize these perspectives and applications.Source: §1.6 · pp. 14–15

Information Source for Explanations

• Explanations can present relevant users or items, charts, sentences, images, or reasoning rules, and multiple explanations may support the same recommendation.The survey organizes explanation research by information source and display style, illustrated by textual, topical, visual, social, statistical, and analytic examples.Source: Information Source for Explanations · pp. 16–17

2.1 Relevant User or Item Explanation

• Relevant-user explanations ground recommendations in similar users’ positive ratings, whereas relevant-item explanations cite previously liked items as similar evidence.User-based displays include aggregated rating histograms or detailed neighbor ratings, while item-based displays show highly rated prior items.Source: §2.1 · pp. 18–19, Figure 2.2 · p. 19
• A user study found that appropriate explanations benefit transparency, scrutability, trustworthiness, effectiveness, persuasiveness, efficiency, and satisfaction.Relevant-item explanations may feel more intuitive and credible, while relevant-user explanations can be less convincing and raise privacy concerns.Source: §2.1 · p. 19, §2.2 · p. 20

2.2. Feature-based Explanation

• Feature-based explanations match user profiles to item attributes such as genres, actors, directors, price, or authors, presenting the features that triggered a recommendation.These explanations are closely related to content-based recommendation and can also use demographic features such as age, gender, and location.Source: §2.2 · p. 20, §2.3 · p. 22

2.2 Feature-based Explanation

• Feature-based explanations improved recommendation effectiveness, while another user study linked explanations closely to users’ trust and satisfaction.Feature information can be displayed through tags, aspect word clouds, or radar charts.Source: §2.2 · pp. 20–21, Figure 2.4 · p. 21, Figure 2.5 · p. 21

2.3. Opinion-based Explanation

• Opinion-based explanations use user-generated reviews and posts to provide finer-grained evidence, displayed either as aspect–sentiment pairs or as explanatory sentences.Aspect-level methods extract or learn aspects and pair them with consumer opinions, while sentence-level methods directly present an explanation sentence.Source: §2.3 · pp. 22–23

2.4. Sentence Explanation

• Sentence explanations are either template-based, which personalize predefined wording and can include dis-recommendations, or generation-based, which produce sentences directly from review data.Template-based recommendation and dis-recommendation explanations improved persuasiveness and conversion rate, while review-based generation must address noisy training text.Source: §2.4 · pp. 24–27

2.5 Visual Explanation

• Visual explanations use attention mechanisms to highlight image regions that influence recommendations, personalizing the highlighted regions to each user’s interests.Lin et al. (2019) applied mutual attention to outfit images, while Chen et al. (2019b) combined image and review information for personalized product-region highlights.Source: §2.5 · pp. 27–28, Figure 2.9 · p. 28, Figure 2.10 · p. 28

2.5. Visual Explanation

• Visual explanations remain at an initial research stage, with advancing image processing expected to improve recommendation performance and explainability.Source: §2.5 · p. 29

2.6 Social Explanation

• Social explanations use friends or similar users to address trust and privacy concerns, influencing recommendation engagement without necessarily predicting actual ratings.Social explanations affected users’ likelihood of checking recommended artists but showed little correlation with their actual artist ratings; social factorization also supports explainable serendipity.Source: §2.6 · pp. 29–30
• Social explanations extend across friend, product, and group recommendation by using social preferences, increasing acceptance, satisfaction, and decision-making efficiency.Methods include graph-based recommendations from similar friends and persuasive-user selection that considers persuasiveness, credibility, and reading cost.Source: §2.6 · pp. 30–31, Figure 2.12 · p. 30

2.7 Summary

• The section categorizes explanations as relevant-user/item, feature, opinion, textual, visual, or social, distinguishing algorithm-linked sources from presentation-focused forms.Relevant-user/item, feature, and opinion explanations are usually tied to recommendation algorithms, whereas textual, visual, and social explanations emphasize how explanations are shown.Source: §2.7 · p. 31

Explainable Recommendation Models

• Explainable recommendation research targets either the explainability of recommendation methods or the explainability of recommendation results.Method-focused research seeks interpretable models, while result-focused research explains outputs without requiring the underlying model to be interpretable.Source: Explainable Recommendation Models · p. 32
• Intrinsic approaches design interpretable models that directly support transparency and explain how the recommendation process works.This philosophy underlies factorization-based, topic-modeling, graph-based, deep-learning, knowledge-based, and rule-mining approaches.Source: Explainable Recommendation Models · p. 32
• Post-hoc/model-agnostic approaches treat the recommendation model as a black box and build separate models to explain its results.This philosophy focuses on explaining recommendation outputs rather than making the recommendation model itself interpretable.Source: Explainable Recommendation Models · p. 32
• The section first supplies machine-learning background for personalized recommendation before surveying explainable recommendation models.The overview is intended to provide readers with necessary background knowledge for the subsequent survey.Source: Explainable Recommendation Models · p. 33

3.1 Overview of Machine Learning for Recommendation

• Latent Factor Models based on Matrix Factorization learn latent factors to predict missing ratings in user-item rating matrices.Representative methods include SVD, NMF (Lee and Seung, 1999, 2001), MMMF, PMF, and LMF (Zhang et al., 2013a,b).Source: §3.1 · p. 33
• Pair-wise and list-wise learning-to-rank methods learn item rankings from implicit feedback, including BPR’s comparison of purchased positive items with unpurchased negative items.Rendle and Schmidt-Thieme (2010) extended pairwise learning to tensor factorization, while Shi et al. (2010) used list-wise learning for collaborative filtering.Source: §3.1 · p. 33
• Deep learning and representation learning expanded recommendation research through models such as Wide and Deep Network, which combines shallow regression with a multilayer neural network.The cited model combines shallow regression and multiple neural-network layers for recommender systems.Source: §3.1 · p. 33

3.2 Factorization Models for Explainable Recommendation

• Latent factor models are difficult to explain because user and item embedding dimensions lack explicitly known meanings.This opacity makes predictions and recommendations challenging to interpret.Source: §3.2 · p. 34
• Explicit Factor Models align latent dimensions with product features extracted from reviews, enabling traceable recommendations and personalized feature-based explanations.They can also explain dis-recommendations and model changing feature preferences dynamically over time.Source: §3.2 · pp. 34–35, Figure 3.1 · p. 35
• Tensor factorization ranks user preferences over items and features across product categories, helping personalize recommendations while alleviating data sparsity.The user-item-feature cube is constructed from product features extracted from textual reviews.Source: §3.2 · pp. 35–36, Figure 3.2 · p. 36
• Multi-task tensor factorization jointly models user preferences for recommendation and opinionated content for feature-level textual explanations.The framework predicts both item preferences and how a user would appreciate an item at the feature level.Source: §3.2 · p. 36, Figure 3.2 · p. 36
• Attention-driven Factor Models learn item-specific user attention distributions over features, which also serve as recommendation explanations.This addresses the limitation of assuming one global feature-preference distribution for each user.Source: §3.2 · p. 36
• Influence analysis explains latent-factor predictions by tracing them to training data and producing intuitive neighbor-style explanations.This model-agnostic, post-hoc approach is discussed more fully in the survey’s subsequent section.Source: §3.2 · pp. 36–37
• Sentiment Utility Logistic Models integrate review aspects and user sentiments into matrix factorization to recommend items together with explanatory features.The method can recommend restaurant aspects such as dining time and dishes, while related work also combines ratings and reviews in aspect-based latent factor models.Source: §3.2 · p. 37
• Factorization trees guide latent-factor learning with regression-tree structure, allowing recommendation explanations to trace latent-profile creation along user and item paths.Tree regularization progressively refines latent factors, making their development inspectable.Source: §3.2 · p. 37

Topic Modeling for Explainable Recommendation

• Figure 3.3 distinguishes explainable recommendation with external data from approaches without external data support, including Abdollahi and Nasraoui (2016).The no-external-data approach is based solely on the user–item rating matrix.Source: Figure 3.3 · p. 38, §3.3 · p. 38
• Explainable Matrix Factorization (EMF) adds an explainability regularizer that aligns user and item latent vectors when neighboring users also purchased the item, producing relevant-user explanations while maintaining rating-prediction accuracy.The explanations frame recommendations as items purchased by users similar to the target user.Source: §3.3 · p. 38
• Liu et al. (2019) model non-randomly missing ratings with influence scores that identify related users, items, and five representative groups for explaining predicted ratings.The groups are influential users, impressionable users, independent users, popular items, and long-tail items.Source: §3.3 · p. 38

3.3 Topic Modeling for Explainable Recommendation

• Topic-modeling approaches widely use textual reviews to generate intuitive explanations, especially topical word clouds for recommender-system users.This research direction includes LDA-based and other probabilistic graphical models for explainable recommendation.Source: §3.3 · pp. 38–39
• HFT links latent-factor dimensions with LDA topics, improving rating prediction while revealing the topics a user most strongly likes for a target item.User latent vectors are projected into latent-topic space to explain particular ratings.Source: §3.3 · p. 39
• Tan et al. (2016) jointly embed users and items in a topical recommendability space, deriving explanations from topic words with the highest recommendability scores.The model integrates recommendability and preference distributions within latent factorization for rating prediction and top-n recommendation.Source: §3.3 · pp. 39–40, Figure 3.4 · p. 39
• Cheng et al. (2019b) combine textual reviews and item images in a multimodal aspect-aware topic model that estimates aspect importance for explainable recommendations.The model learns user preferences and item features from different modalities and aspects.Source: §3.3 · p. 40
• Wu and Ester (2015) propose FLAME to learn personalized sentiment toward item aspects, improve TripAdvisor hotel recommendation performance, and explain recommendations with sentiment-scaled aspect word clouds.FLAME combines collaborative filtering with aspect-based opinion mining and predicts aspect ratings for new items.Source: §3.3 · pp. 40–41, Figure 3.5 · p. 40
• Zhao et al. (2015) integrate sentiment, aspect, and region information for explainable POI recommendation, using each user’s topical-aspect preferences and cluster sentiments as explanations.The sentiment attached to each topical cluster distinguishes this approach from explanations based only on topical clusters.Source: §3.3 · p. 41
• Ren et al. (2017) model social explainable recommendations with sCVR, representing viewpoints as concept-topic-sentiment tuples from reviews and trusted social relations.Viewpoints improve prediction accuracy and generate explanations from the user’s favorite embedded topics.Source: Figure 3.4 · p. 39, §3.3 · p. 41

3.4 Graph-based Models for Explainable Recommendation

• TriRank models user–item–aspect ternary relations in a tripartite graph, ranking vertices with smoothness and fitting constraints for personalized top-N recommendation explanations.Explanations use the highest-ranked aspects matching the target user and recommended item.Source: §3.4 · p. 41, Figure 3.6 · p. 42
• OCuLaR applies overlapping co-clustering to a user–item bipartite graph without external aspects, identifying groups of users with similar interests and items with similar properties.The clustering structure supports explanations based on collaborative information.Source: §3.4 · pp. 41–42, Figure 3.7 · p. 42
• Co-cluster explanations can state that a user receives an item because of prior purchases and because similarly profiled users also purchased it.When a user–item pair belongs to multiple co-clusters, the method can generate multiple user-based and item-based explanations.Source: §3.4 · p. 42, Figure 3.7 · p. 42
• Wang et al. (2018c) combine tree-based decision rules with embeddings to retain rule-based explainability while generalizing to unseen user–item cross features.Rules come from side-information cross features, while collaborative learning supports embedding-based generalization.Source: §3.4 · p. 43
• An attention network in Wang et al. (2018c) identifies the most significant decision rules used during recommendation, providing the model’s explanations.The decision rules are learned first by a tree-based model before being incorporated into the embedding model.Source: §3.4 · p. 43
• Graph-based explainable recommendation is especially suited to social recommendation because social networks naturally form graphs.The survey therefore includes social-network methods alongside user–item graph models.Source: §3.4 · p. 43
• UniWalk (Park et al., 2018) combines ratings with social-network structure, explaining product recommendations through friends who have similar preferences.The explanation is derived from the target user’s graph-based social connections.Source: §3.4 · p. 43

• Deep explainable recommendation applies CNNs, RNNs/LSTMs, attention, memory, and capsule networks to rating, top-n, and sequential recommendation tasks.Natural-language generation can replace predefined explanation templates with automatically generated explanation sentences.Source: §3.5 · p. 43
• Attention mechanisms explain recommendations by highlighting important review words, selecting useful reviews, or identifying influential items in a user’s history.These approaches support rating prediction, faster product evaluation, and sequential recommendation explanations based on prior interactions.Source: §3.5 · pp. 44, 47, Figure 3.8 · p. 44
• Explainable deep models can use explicit feature hierarchies or learned disentangled features to improve explainability, user control, and practical usability.DEAML retrieves multi-level features from explicit hierarchies, while disentangled representations enable users to better control recommendation results.Source: §3.5 · p. 45
• Natural-language models generate personalized explanations from ratings, feature words, review summaries, or jointly learned recommendation and explanation representations.Methods include character-level RNNs, tip generation, topic-sensitive generation, and co-attentive encoder-selector-decoder architectures; crowdsourcing can also synthesize personalized explanations from review quotes [Costa et al., 2018; Li et al., 2017; Chen et al., 2019a; Chen et al., 2019d; Chang et al., 2016].Source: §3.5 · pp. 45–46, Figure 3.9 · p. 45
• Other approaches explain recommendations through highlighted image regions, memory-slot reasoning over interaction histories, or sentiment-bearing capsule logic units.These mechanisms connect visual regions, prior items, user intents, or item-aspect–user-viewpoint pairs to recommendation decisions [Chen et al., 2019b; Chen et al., 2018c; Tao et al., 2019b; Li et al., 2019].Source: §3.5 · pp. 47–48
• The fidelity of explanations from complex deep models remains unresolved, particularly because attention may not faithfully reflect the mechanisms producing recommendations.Jain and Wallace (2019) questioned attention’s explanatory meaningfulness, whereas Wiegreffe and Pinter (2019) argued that conclusions depend on the definition of explanation; more advanced models are needed [Jain and Wallace, 2019; Wiegreffe and Pinter, 2019].Source: §3.5 · pp. 48–49

• Knowledge graphs encode rich user and item information that supports intuitive, tailored explanations for recommended items.Source: §3.6 · p. 49
• Catherine et al. (2017) jointly ranked movies and knowledge graph entities with Personalized PageRank, producing recommendations alongside ranked entity explanations.Source: §3.6 · p. 49
• Ai et al. (2018) used knowledge graph embeddings for explainable recommendation by constructing a user–item knowledge graph.Source: §3.6 · p. 49

• Ripple Network propagates preferences from historically clicked items through KG links, combining activated ripples to predict candidate-item click probabilities and provide path-based explanations (Wang et al., 2018a).Source: §3.6 · p. 50
• Sequential KG recommenders combine RNN-based sequential preferences with memory-based attribute preferences, explaining recommendations through influential attributes and their important values.Source: §3.6 · pp. 50–51
• Recommendations can be explained by finding shortest or navigated paths between users and recommended items in the knowledge graph, including policy-guided reinforcement reasoning.Path-based reasoning avoids enumerating all user–item paths, which is computationally prohibitive in huge real-world knowledge graphs.Source: §3.6 · pp. 50–51, Figure 3.11 · p. 51
• Multimodal KG modeling captures users’ dynamic preferences at the interaction level by representing sequential user–item interactions over knowledge graphs (Huang et al., 2019).Source: §3.6 · p. 51
• KG inductive rules connect recommended items with users’ purchase histories through human-readable multi-hop relational patterns, while jointly guiding neural recommendation.Source: §3.6 · p. 51

3.7. Rule Mining for Explainable Recommendation

• A reinforcement-learning agent can learn to navigate from users to correct items and then directly walk to recommended items at inference time without enumerating all user-item paths.The approach trains the agent with high rewards for reaching correct items.Source: §3.7 · p. 52

3.7 Rule Mining for Explainable Recommendation

• Knowledge-graph differentiable paths can explain recommendations from embedding-based black-box models while jointly regularizing those models with structured knowledge for better performance (Zhang et al., 2020).The approach uses end-to-end joint learning to combine embedding-based and path-based recommendation models.Source: §3.7 · p. 52
• Rule mining approaches are valuable for explainable recommendation because they can generate straightforward explanations for users, with association rule mining being the most frequently used technique (Agrawal et al., 1993, 1994).The classic “beer-diaper” example originated in data mining research.Source: §3.7 · p. 52
• Association and frequent-pattern mining have been applied to large-scale web-page, web-shop, item-similarity, conversational, robust collaborative, and task-based recommendation systems (Mobasher et al., 2001; Cho et al., 2002; Smyth et al., 2005; Sandvig et al., 2007; Zhang et al., 2015a).Amatriain and Pujol (2015) surveyed data mining for personalized recommendation systems.Source: §3.7 · pp. 52–53
• Personalized association rule mining extracts rules for target users, making recommendations self-explainable through associations such as shared user preferences (Lin et al., 2000, 2002).An example explanation is that 90% of articles liked by users A and B are also liked by user C.Source: §3.7 · p. 53
• YouTube used association rules over co-watched video pairs and user-specific seed videos to recommend related videos, while the seed video and triggering rules served as explanations.The system formed seed sets from recently watched and explicitly favorited, liked, rated, or playlist-added videos, and Figure 3.12 shows the resulting personalized recommendation explanations.Source: §3.7 · p. 53, Figure 3.12 · p. 53

3.8. Model Agnostic and Post Hoc Explainable Recommendation

• Balog et al. (2019) infer tag-based preferences from rated items to produce transparent recommendations with sentence-level explanations and user feedback control.Users can scrutinize individual sentences, and preference changes immediately affect recommendations.Source: §3.8 · p. 54

3.8 Model Agnostic and Post Hoc Explainable Recommendation

• Post-hoc methods generate explanations independently from the recommendation mechanism, so explanations can remain truthful while being decoupled from how recommendations were produced.Templates may be selected using statistics such as maximum confidence, but post-hoc explanations need not be fake.Source: §3.8 · pp. 54–55
• Peake and Wang (2018) extract association rules from black-box recommendation transactions, while Singh and Anand (2018) train tree-based surrogates to explain ranking lists.The former is point-wise, whereas the latter is pair-wise; Peake and Wang also evaluate what percentage of recommendations can be explained using fidelity.Source: §3.8 · p. 55

3.8. Model Agnostic and Post Hoc Explainable Recommendation

• LIME-style local surrogate explanations approximate complex models around individual samples, and Singh and Anand (2019) apply this idea to ranking with linear SVM explanations.The ranking problem is converted into binary relevance classification, with coefficients indicating words associated with document relevance.Source: §3.8 · pp. 55–56
• Wang et al. (2018d) use reinforcement-learning agents to generate sentence explanations for any recommendation model by rewarding prediction accuracy and presentation quality.The recommendation model is treated as a black-box environment, and the agents learn explanations through expected reward optimization.Source: Figure 3.13 · p. 56, §3.8 · pp. 56–57
• McInerney et al. (2018) use bandits to jointly learn user-specific explanation responses, recommendation content, and exploration-exploitation trade-offs, outperforming the best static explanation ordering.The experiments show that explanations affect users’ responses to recommendations.Source: §3.8 · p. 56
• Cheng et al. (2019a) introduce FIA, which uses influence functions to trace latent-factor predictions to historical user-item interactions and generate neighbor-style explanations.The method measures how training interactions affect predictions in latent factor models.Source: §3.8 · p. 57

3.8 Model Agnostic and Post Hoc Explainable Recommendation

• Overall, post-hoc explanations offer flexibility across recommendation models, but their fidelity may be limited because they need not follow the exact recommendation mechanism.The explanation model is developed separately to explain a black-box prediction model.Source: §3.8 · p. 57

3.9 Summary

• The section surveys machine-learning models used for explainable recommendation.It first reviews machine learning for recommender systems before focusing on explainable recommendation methods.Source: §3.9 · p. 57
• The surveyed approaches include matrix/tensor factorization, topic modeling, graph-based, and deep learning methods.Source: §3.9 · p. 57
• The survey also covers knowledge-based, rule mining, and post-hoc/model-agnostic explainable recommendation methods.Source: §3.9 · p. 57

3.9. Summary

• Explainability can concern either recommendation methods or recommendation results, corresponding to model-intrinsic and model-agnostic approaches.Source: §3.9 · p. 58
• Both modeling philosophies have roots in human cognitive science.Source: §3.9 · p. 58
• Transparent models explain decisions through step-by-step reasoning, whereas post-hoc methods explain intuitive decisions after they are made.Source: §3.9 · p. 58
• The section concludes that determining the correct philosophy requires advances in human cognitive science and understanding of the human brain.It suggests that neither approach can currently be established as definitively correct, and possibly both may be appropriate.Source: §3.9 · p. 58

Evaluation of Explainable Recommendation

• Explainable recommendation evaluation should assess both recommendation performance and explanation quality, using conventional recommendation metrics alongside explanation-focused online and offline methods.Rating prediction may use MAE or RMSE, top-n recommendation may use precision, recall, F-measure, or NDCG, and online ranking may use CTR or conversion rate.Source: Evaluation of Explainable Recommendation · p. 59, §4.5 · p. 67

4.1 User Study

• User studies use surveys or tasks with recruited participants to assess explanation effectiveness, justification, understandability, mood compatibility, trust, and informativeness.Studies have evaluated explanation styles, interfaces, textual explanations, visual explanations, and viewpoint-based explanations through crowdsourcing, volunteers, paid subjects, or manual labeling.Source: §4.1 · pp. 60–61

4.2 Online Evaluation

• Online experiments primarily evaluate explanation persuasiveness, effectiveness, efficiency, and satisfaction through observed user behavior, while A/B tests compare personalized, baseline, and no-explanation conditions.Online metrics depend on available system information and may include CTR, conversion or purchase rate, agreement, and gross profit; user studies permit deeper surveys but can introduce cognitive bias.Source: §4.2 · pp. 62–63, §4.3 · p. 64

4.3 Offline Evaluation

• Offline evaluation measures either how many recommendations are explainable or the quality of explanations, using metrics such as MEP, MER, model Fidelity, text-generation scores, readability indices, and ground-truth matching.Offline evaluation is easier to implement than online evaluation but remains limited by the availability of ground-truth explanations and the need for more comprehensive explainability protocols.Source: §4.3 · pp. 64–65

4.4. Qualitative Evaluation by Case Study

• Case studies provide qualitative insight into why recommendations work or fail, including sequential “one-to-multiple” and “one-to-one” behavior patterns and aspect-based explanations for hotel recommendations.For hotel recommendations, comparing user preferences with item quality explains why one item satisfies the target user better than another.Source: §4.4 · pp. 65–66, Figure 4.3 · p. 66, Figure 4.4 · p. 66

5.1 Explainable E-commerce Recommendation

• E-commerce is a standard and widely adopted setting for explainable recommendation, spanning aspect, review, visual, sequential, cross-category, and product-search methods that help users make persuasive, efficient, informed, and socially responsible choices.Representative systems have been evaluated on commercial websites and Amazon data using graph ranking, learning-to-rank, attention, co-clustering, memory networks, tensor factorization, and explicit queries.Source: §5.1 · pp. 68–69

5.2. Explainable Point-of-Interest Recommendation

• Point-of-interest explanations help users save time and avoid costly travel mistakes, plan routes, and accept recommendations, with methods extracting aspects, sentiments, regions, and user-aspect relationships from review and location data.Reported applications include restaurants, hotels, beauty and spa services, tourist locations, and location-based social networks.Source: §5.2 · p. 70

5.3 Explainable Social Recommendation

• Social explainable recommendation supports trust, transparency, efficiency, and credibility by exposing social connections, information sources, group realities, or cross-references that clarify friend, news, and other social recommendations.Such explanations can help users distinguish credible from fake news, understand why recommendations matter, and improve perceptions of group recommendations.Source: §5.3 · p. 71, §5.4 · p. 72

5.4 Explainable Multimedia Recommendation

• Explainable multimedia recommendation spans books, news, music, movies, and videos, helping users decide efficiently while reducing unnecessary data-intensive media transmissions.Approaches include explainable matrix factorization, crowd-sourced and story-based explanations, and knowledge-graph reasoning over movie entities.Source: §5.4 · pp. 72–73
• Personalized natural-language reasons for song recommendations significantly outperform manually selected reasons in click-through rate in large-scale online environments (Zhao et al., 2019a).Multimedia systems also explain recommendations through related-video association rules, political topics, and knowledge paths connecting users to recommended items.Source: §5.4 · p. 73

5.5 Other Explainable Recommendation Applications

• Explainability is expanding to academic, citation, legal, and healthcare recommendation, where interpretable decision rules and outlier detection help clarify data-driven decisions.Healthcare examples classify forum sentences using extracted decision rules and study interpretable outlier detection for health monitoring.Source: §5.5 · p. 73, §5.6 · p. 74

5.6. Summary

• Explanations can hurt in time-critical or emergency scenarios, so systems should avoid unnecessary detail, repetition, obvious explanations, and misleading or incorrect rationales.Examples include driving, emergency medical decisions, and battlefield decisions, where users may not have time to evaluate explanations.Source: §5.6 · p. 74

6.1.1 Explainable Deep Learning for Recommendation

• Explainable deep recommendation remains immature because both explanation-generating components and the underlying recommendation models are often black boxes with unintuitive latent representations.Future work should clarify latent component meanings and interactions, benefiting recommendation and applications in vision, language, healthcare, education, chatbots, and autonomous systems.Source: §6.1.1 · p. 75, §6.1 · p. 76

6.1.2 Knowledge-enhanced Explainable Recommendation

• Knowledge-enhanced models can produce tailored, high-fidelity explanations through domain reasoning, while multimodal research should align heterogeneous sources, transfer explanations across domains, and study effects on user receptiveness.These directions extend beyond unstructured text and images toward knowledge graphs, conversational recommendation, and multimodal evidence.Source: §6.1.2 · pp. 76–77, §6.1.3 · p. 77

6.1.4 Context-aware Explanations

• Context-aware explainable recommendation should adapt explanations to changing user preferences, time, and location, since most current models profile users statically from training data.The survey identifies extensive exploration as necessary for context-aware explanations.Source: §6.1.4 · p. 77

6.1.5 Aggregation of Different Explanations

• Future systems must select and aggregate diverse candidate explanations into logically consistent outputs, integrating statistical and logical reasoning rather than presenting potentially conflicting or unnecessarily complete explanations.The survey connects this challenge to user satisfaction, cognitive science, and logical inference in decision-making systems.Source: §6.1.5 · p. 78

6.1.6 Explainable Recommendation as Reasoning

• Explainable recommendation should combine symbolic and machine-learning approaches, including causal and neural logic reasoning, to advance collaborative filtering from latent similarity matching toward collaborative reasoning.Natural-language explanations are another open direction: current review-like generation remains early, noisy, and insufficiently personalized, while future systems should incorporate language models, images, knowledge graphs, and sentiment.Source: §6.1.6 · pp. 78–79, §6.1.7 · p. 79

6.1.8 Answering the “Why” in Conversations

• Explainable recommendation should support conversational systems in which users actively ask why an item was recommended, extending explanation beyond passive display and broadening the queries intelligent systems can process.The section also calls for user-behavior research on interaction and receptiveness, which can improve evaluation of explanations.Source: §6.1.8 · p. 80

6.2.1 Evaluation of Explainable Recommendations

• Evaluation needs reliable offline explainability measures alongside traditional ranking metrics and user studies, covering user effects such as persuasiveness, transparency, trustworthiness, satisfaction, and algorithmic fidelity.Different explanation forms may require different evaluation methods, while user behavior remains essential because recommender systems are human-computer interaction systems.Source: §6.2.1 · pp. 80–81

6.3 Explanation for Broader Impacts

• Broader research should examine explanations’ effects on trustworthiness, efficiency, diversity, satisfaction, scrutability, and fairness, while cognitive science must clarify whether transparent models, post-hoc explanations, or both best reflect human decision-making.The relationship between transparency and fairness in information systems remains unresolved, and answering the broader methodological question requires advances in understanding human cognition and the brain.Source: §6.3 · p. 81, §6.4 · p. 82

• Complex models improve recommendation performance but increase the difficulty of transparency and explainability.Source: Conclusions · pp. 83–84
• Explainability is needed both for users to understand recommendation results and for designers to understand recommendation mechanisms.Without explanations, user persuasion and trustworthiness may decrease, while decision-support systems require users to understand results and act on them.Source: Conclusions · pp. 83–84
• The survey reviews explanation types, explainable recommendation models, evaluation methods, and applications spanning e-commerce, POI, social, and multimedia recommendations.Covered model families include MF-based, topic-based, graph-based, deep learning-based, knowledge-based, mining-based, and post-hoc approaches.Source: Conclusions · p. 84
• Future progress may draw on knowledge graphs, deep learning, natural language processing, user behavior analysis, model aggregation, logical reasoning, conversational systems, and cognitive foundations.Source: Conclusions · p. 84
• The broader rise of Explainable AI underscores the importance of addressing explainability across search and recommendation systems.Relevant AI areas include deep learning, computer vision, autonomous driving, and natural language processing.Source: Conclusions · p. 84