ExplainableRecommendation:ASurveyand …

1 Explainable Recommendation: ASurvey and New PerspectivesYongfeng Zhang1and Xu Chen21 Rutgers University, USA; University, China; recommendation attempts to develop modelsthat generate not only high-quality recommendations butalso intuitive explanations. The explanations may either bepost-hoc or directly come from an explainable model (alsocalled interpretable or transparent model in some contexts).Explainable recommendation tries to address the problemofwhy: by providing explanations to users or system design-ers, it helps humans to understand why certain items arerecommended by the algorithm, where the human can eitherbe users or system designers. Explainable recommendationhelps to improve the transparency, persuasiveness, effective-ness, trustworthiness, and satisfaction of recommendationsystems.

2 It also facilitates system designers for better systemdebugging. In recent years, a large number of explainable rec-ommendation approaches especially model-based methods have been proposed and applied in real-world this survey, we provide a comprehensive review for theexplainable recommendation research. We first highlight theposition of explainable recommendation in recommendersystem research by categorizing recommendation problemsinto the 5W, , what, when, who, where, and why. We thenYongfeng Zhang and Xu Chen (2020), Explainable Recommendation: A Surveyand New Perspectives , Foundations and Trends in Information Retrieval: Vol. 14,No. 1, pp 1 101. DOI: version of record is available at: a comprehensive survey of explainable recommen-dation on three perspectives: 1) We provide a chronologicalresearch timeline of explainable recommendation, includinguser study approaches in the early years and more recentmodel-based approaches.

3 2) We provide a two-dimensionaltaxonomy to classify existing explainable recommendationresearch: one dimension is the information source (or displaystyle) of the explanations, and the other dimension is thealgorithmic mechanism to generate explainable recommen-dations. 3) We summarize how explainable recommendationapplies to different recommendation tasks, such as productrecommendation, social recommendation, and POI also devote a section to discuss the explanation per-spectives in broader IR and AI/ML research. We end thesurvey by discussing potential future directions to promotethe explainable recommendation research area and version of record is available at: Explainable RecommendationExplainable recommendation refers to personalized recommendationalgorithms that address the problem ofwhy they not only provide usersor system designers with recommendation results, but also explanationsto clarify why such items are recommended.

4 In this way, it helps toimprove the transparency, persuasiveness, effectiveness, trustworthiness,and user satisfaction of the recommendation systems. It also facilitatessystem designers to diagnose, debug, and refine the highlight the position of explainable recommendation in therecommender system research area, we classify personalized recommen-dation with a broad conceptual taxonomy. Specifically, personalizedrecommendation research can be classified into the 5W problems when, where, who, what, and why, corresponding to time-aware rec-ommendation (when), location-based recommendation (where), socialrecommendation (who), application-aware recommendation (what), andexplainable recommendation (why), where explainable recommendationaims to answerwhy-type questions in recommender version of record is available at: recommendation models can either be model-intrinsicor model-agnostic (Lipton,2018; Molnar,2019).

5 The model-intrinsicapproach develops interpretable models, whose decision mechanism istransparent, and thus, we can naturally provide explanations for themodel decisions (Zhanget al., 2014a). The model-agnostic approach(Wanget al.,2018d), or sometimes called the post-hoc explanationapproach (Peake and Wang,2018), allows the decision mechanism tobe a blackbox. Instead, it develops an explanation model to generateexplanations after a decision has been made. The philosophy of these twoapproaches is deeply rooted in our understanding of human cognitivepsychology sometimes we make decisions by careful, rational reasoningand we can explain why we make certain decisions; other times we makedecisions first and then find explanations for the decisions to supportor justify ourselves (Lipton,2018; Miller,2019).

6 The scope of explainable recommendation not only includes de-veloping transparent machine learning, information retrieval, or datamining models. It also includes developing effective methods to deliverthe recommendations or explanations to users or system designers, be-cause explainable recommendations naturally involve humans in theloop. Significant research efforts in user behavior analysis and human-computer interaction community aim to understand how users interactwith this section, we will introduce not only the explainable recom-mendation problem, but also a big picture of the recommender systemresearch area. It will help readers to understand what is unique aboutthe explainable recommendation problem, what is the position of ex-plainable recommendation in the research area, and why explainablerecommendation is important to the A Historical OverviewIn this section, we will provide a historical overview of the explainablerecommendation research.

7 Though the termexplainable recommendationwas formally introduced in recent years (Zhanget al., 2014a), the basicconcept, however, dates back to some of the earliest works in personalizedThe version of record is available at: A Historical Overview5recommendation research. For example, Schaferet al.(1999) noted thatrecommendations could be explained by other items that the user isfamiliar with, such asthis product you are looking at is similar to theseother products you liked before, which leads to the fundamental idea ofitem-based collaborative filtering (CF); Herlockeret al.(2000) studiedhow to explain CF algorithms in MovieLens based on user surveys; andSinha and Swearingen (2002) highlighted the role of transparency inrecommender systems.

8 Besides, even before explainable recommendationhas attracted serious research attention, the industry has been usingmanual or semi-automatic explanations in practical systems, such asthepeople also viewedexplanation in e-commerce systems (Tintarevand Masthoff,2007a).To help the readers understand the pre-history research of recom-mendation explanation and how explainable recommendation emergedas an essential research task in the recent years, we provide a historicaloverview of the research line in this approaches to personalized recommender systems mostly fo-cused on content-based or collaborative filtering (CF)-based recommen-dation (Ricciet al.,2011). Content-based recommender systems modeluser and item profiles with various available content information, suchas the price, color, brand of the goods in e-commerce, or the genre,director, duration of the movies in review systems (Balabanovi andShoham,1997; Pazzani and Billsus,2007).

9 Because the item contentsare easily understandable to users, it was usually intuitive to explainto users why an item is recommended. For example, one straightfor-ward way is to let users know the content features he/she might beinterested in the recommended item. Ferwerdaet al.(2012) provided acomprehensive study of possible protocols to provide explanations forcontent-based , collecting content information in different application do-mains is time-consuming. Collaborative filtering (CF)-based approaches(Ekstrandet al.,2011), on the other hand, attempt to avoid this dif-ficulty by leveragingthe wisdom of crowds. One of the earliest CFalgorithms is User-based CF for the GroupLens news recommendationsystem (Resnicket al.)

10 ,1994). User-based CF represents each user asa vector of ratings, and predicts the user s missing rating on a newsThe version of record is available at: based on the weighted average of other users ratings on themessage. Symmetrically, Sarwaret al.(2001) introduced the Item-basedCF method, and Lindenet al.(2003) further described its applicationin Amazon product recommendation system. Item-based CF takes eachitem as a vector of ratings, and predicts the missing rating based onthe weighted average of ratings from similar the rating prediction mechanism would be relatively difficultto understand for average users, user- and item-based CF are somewhatexplainable due to the philosophy of their algorithm design.