Explainable recommendation A survey and new perspective _第二章_个人笔记

2 Information Source for Explanations

前文 1.3 Classification of the Methods提出可解释推荐两个维度的分类，本章从第一个维度——information source or display style，展开介绍。

推荐可解释性可能来自不同的信息源并且以不同的形式呈现。如，相关的user或item，雷达图，句子，图像或者一些推理规则。另外，同样的推荐可能有多种解释。

Recommendation explanations can be generated from different information sources and be presented in different display styles (Tintarev and Masthoff, 2015), e.g.,a relevant user or item, a radar chart, a sentence, an image, or a set ofreasoning rules. Besides, there could exist many different explanationsfor the same recommendation.Besides, there could exist many different explanations for the same recommendation.

一些举例 ==

For example, Zhang et al. (2014a) generated (personalized) textual sentences as explanations to help users understand each recommendation; Wu and Ester (2015), Zhang (2015), and Al-Taie and Kadry (2014) provided topical word clouds to highlight the key features of a recommended item; Chen et al. (2019b) proposed visually explainable recommendations, where certain regions of a product image are highlighted as the visual explanations; Sharma and Cosley (2013) and Quijano-Sanchez et al. (2017) generated a list of social friends who also liked the recommended product as social explanations. In early research stages, Herlocker et al. (2000), Bilgic and Mooney (2005), and Tintarev and Masthoff (2007b) and McSherry (2005) adopted statistical histograms or pie charts as explanations to help users understand the rating distribution and the pros/cons of a recommendation. Duet al. (2019) provided a visual analytics approach to explainable recommendation for event sequences. Figure 2.1 shows several representative recommendation explanations.

2.1 Relevant User or Item Explanation （基于协同过滤）

本节主要介绍最早期使用的user-based CF和 item-based CF，及其拓展。

In this section, we introduce explainable recommendation based on user- and item-based collaborative filtering (Cleger-Tamayo et al., 2012; Resnick et al., 1994; Sarwar et al., 2001; Zanker and Ninaus, 2010) – two fundamental methods for personalized recommendation.Extensions of the two basic methods will also be introduced in this section.

这里两种CF都是利用用户隐式或显式的反馈信息。

User-based and item-based explanations are usually provided based on users’ implicit or explicit feedback

1. user-based CF

在基于用户的协同过滤（Resnick et al.，1994）中，我们首先为目标用户找到一组相似的用户（即邻居）。当算法向目标用户推荐一个项目，推荐的解释就是该用户与一组“邻居”用户相似，这些邻居用户对推荐的item评价都很好。

In user-based collaborative filtering (Resnick et al., 1994), we first find a set of similar users (i.e.,neighbors) for the target user. Once the algorithm recommends an item to the target user, the explanation is that the user is similar to a group of “neighborhood” users, and these neighborhood users made good ratings on the recommended item.

例子：

For example, Herlocker et al. (2000)compared the effectiveness of different display styles for explanations in user-based collaborative filtering. In this research, explanations can be displayed as an aggregated histogram of the neighbors’ ratings, or be displayed as the detailed ratings of the neighbors, as shown in Figure 2.2.Recent model-based explainable recommendation approaches can generate more personalized and meticulously designed explanations than this,but this research illustrated the basic ideas of providing explanations in recommender systems.

2. item-based CF

在基于项目的协同过滤（Sarwar et al.，2001）中，可以通过告诉用户推荐的项目与用户之前喜欢的其他项目相似来提供解释，如图2.3的左边所示的与之相似的，用户的几部高评级电影（4星或5星）。

更直观地说，如图2.1所示，一个基于user的推荐解释告诉Bob，相似的用户William和Fred也购买了这个产品；而基于item的推荐的解释则通过展示他之前购买的相机来说服Bob购买镜头。

Tintarev(2007)研究可解释性对系统的作用，提出七个方面好处：透明度、严谨性、可信赖性、有效性、说服力、效率、满意度。（ transparency, scrutability, trustworthiness, effectiveness, persuasiveness, efficiency, and satisfaction.）

To study how explanations help in recommender systems, Tintarev(2007) developed a prototype system to study the effect of different types of explanations, especially the relevant-user and relevant-item explanations. In particular, the author proposed seven benefits of providing recommendation explanations, including transparency, scrutability, trustworthiness, effectiveness, persuasiveness, efficiency, and satisfaction.Based on their user study, the author showed that providing appropriate explanations can indeed benefit the recommender system over these seven perspectives.

3. 总结扩展

相比之下item-based CF推荐对用户来说更直接，所以认为是可相信的解释。

Usually, relevant-item explanations are more intuitive for users to understand because users are familiar with the items they interacted before. As a result, these items can serve as credible explanations for users.

而user-based CF 推荐说服力弱，因为用户也不了解自己所在的群体。

另外，因为关联其他用户的信息涉及到隐私问题，所以基于relevant-user的推荐有一个新的发展方向——基于开放社交信息的推荐。

-Relevant-user explanations, however, could be less convincing because the target user may know nothing about other “similar” users at all, which may decrease the trustworthiness of the explanations(Herlocker et al., 2000).
-Besides, disclosing other users’ information may also cause privacy problems in commercial systems. This problem drives relevant-user explanation into a new direction, which leverages social friend information to provide social explanations (Ren et al., 2017;Tsai and Brusilovsky, 2018). For example, we can show a user with her friends’ public interest as explanations for our social recommendations. In the following, we will review this research direction in the social explanation section (Section 2.6)

2.2 Feature-based Explanation(基于内容)

1.什么是 feature-based explanation

基于特征的解释和基于内容的推荐很接近。

The feature-based explanation is closely related to content-based recommendation methods.

在基于内容的推荐中，系统通过匹配用户画像和候选物品的内容特征来进行推荐。

In content-based recommendation, the system provides recommendations by matching the user profile with the content features of candidate items (Cramer et al., 2008; Ferwerda et al., 2012; Pazzani and Billsus, 2007).

基于内容的推荐通常基于特征，来更直观的进行解释。

Content-based recommendations are usually intuitive to explain based on the features.

根据应用情景，基于内容的推荐通过不同的item特征产生。例如，电影推荐领域，通常基于电影类型、演员或导演来推荐。而图书推荐领域，根据图书类型、价格或作者进行推荐。

基于特征的解释的常见范式是给用户呈现那些符合用户画像的特征。

A conventional paradigm for feature-based explanation is to show users with the features that match the user’s profile.

举例：

Vig等人。（2009）采用电影标签作为特征来生成推荐和解释，如图2.4所示。通过展示电影的特征并且告诉用户为什么这些特征和他们相关，来解释推荐的电影。
作者还设计了一个用户研究，表明提供基于特征的解释有助于提高推荐的有效性。此外，Ferwerda等人。（2012）进行了一项用户研究，结果支持了这样一个观点，即解释与用户对建议的信任和满意度高度相关。

-Vig et al. (2009) adopted movie tags as features to generate recommendations and explanations, as shown in Figure 2.4. To explain the recommended movie, the system displays the movie features and tells the user why each feature is relevant to her.
-The authors also designed a user study and showed that providing feature-based explanations can help to improve the effectiveness of recommendations. Furthermore, Ferwerda et al. (2012) conducted a user study, and the results supported the idea that explanations are highly correlated with user trust and satisfaction in the recommendations.

2. feature-based explanation 呈现形式

内容特征可以通过许多不同的解释风格呈现。

Content features can be displayed in many different explanation styles.

例如：Hou et al. (2018) 使用雷达图解释为什么推荐一个item，以及不推荐其他item的原因。如Figure 2.5,推荐的解释是，大部分方面满足目标用户的偏好。
（左图显示酒店推荐，item1被推荐给用户，因为大部分方面满足用户偏好。右图类似，视频游戏推荐，同样item1满足用户大部分偏好。）

For example, Hou et al. (2018) used radar charts to explain why an item is recommended and why others are not. As shown in Figure 2.5, a recommendation is explained in that most of its aspects satisfy the preference of the target user.

3. 一个重要依据信息——user demographic information

用户的人口统计信息(demographic information)描述了用户的内容特征。并且人口统计信息也可以被用来生成基于特征的推荐。
基于人口学的推荐（Pazzani，1999）是最早实现个性化推荐系统的方法之一。最近，研究人员还将人口学方法整合到社交媒体中，在社交环境中提供产品推荐（Zhao et al.，2014，2016）。

-User demographic information describes the content features of users, and the demographic features can also be used to generate feature-based explanations.
-Demographic-based recommendation (Pazzani, 1999) is one of the earliest approaches to personalized recommendation systems. Recently, researchers have also integrated demographic methods into social media to provide product recommendations in social environments (Zhao et al., 2014, 2016).

基于人口统计的方法（The demographic-based approach）根据用户的人口统计特征（如年龄、性别和居住地点）做出推荐。直观地说，根据人口统计信息推荐的商品可以用对应的人口统计特征来解释，例如，告诉用户“在这个年龄段，80%的顾客都买了这个产品”。
zhao等。（2014）在同一人口统计特征空间中代表产品和用户，并使用排序函数学习的特征权重来解释结果；zhao等。（2016）进一步探索社交媒体环境下的人口统计信息，进行产品推荐，并进行基于特征的解释。

-The demographic-based approach makes recommendations based on user demographic features such as age, gender, and residence location. Intuitively, an item recommended based on demographic information can be explained by the demographic feature(s) that triggered the recommendation, e.g., by telling the user that “80% of customers in your age bought this product” .
-Zhao et al. (2014) represented products and users in the same demographic feature space, and used the weights of the features learned by a ranking function to explain the results; Zhao et al. (2016) further explored demographic information in social media environment for product recommendation with feature-based explanations.

2.3 Opinion-based Explanation （基于内容）

1. 数据来源——user-generated contents

用户生成的内容积累的越来越多，例如电子商务评论、社交媒体发言。用户用这些内容表达他们对特定item或aspects的意见

More and more user-generated contents have been accumulating on the Web, such as e-commerce reviews and social media posts, which help users to express their opinion on certain items or aspects.

研究者已经表明，这些信息在用户画像构建和推荐方面非常有用。

Researchers have shown that such information is quite beneficial in user profiling and recommendation (McAuley and Leskovec, 2013; Zheng et al., 2017).

另外，这些信息有助于生成更细粒度（finer-grained）和更可靠的解释。

Besides, it helps to generate finer-grained and more reliable explanations, which benefit users to make more informed decisions (Li et al., 2017; Zhang et al., 2014a).

基于此人们提出许多模型，来解释基于用户生成文本的推荐。

With this motivation, many models have been proposed to explain recommendations based on user-generated texts.

---

2. Display style：aspect-level & sentence level

这一方向的方法，可根据展示的方式，大致分为aspect-level 和 setence-level 方法。

Methods in this direction can be broadly classified into aspect-level and sentence-level approaches, according to how the explanations are displayed.

如图2.1（主要在图右侧）
aspect-level models 呈现item信息如color，quality和它们分数作为解释。

where aspect-level models present item aspects (such as color, quality) and their scores as explanations,

而sentence-level models 直接呈现一个关于为什么推荐这个相机镜头的解释性的句子

while the sentence-level models directly present an explanation sentence to users about why the camera lens is recommended.

本节下面内容主要介绍aspect-levelexplanation，而sentence-level explanations 在下节介绍。

We will focus on aspect-level explanation in this subsection, while sentence-level explanations will be introduced in the following subsection together with other natural language generation-based explanation models.

---

3. 什么是 aspect-level 及其应用

aspect-level explanation 和 feature-based explanation相似，只不过aspects通常不能直接在item 或 user profile里被使用。

The aspect-level explanation is similar to feature-based explanation, except that aspects are usually not directly available in an item or user profile.

但是，aspects 作为推荐系统的一部分，（例如从评论中）被抽取或者学习得到。并且aspects可以和用户opinion成对结合，表达更清晰的情感。

Instead, they are extracted or learned as part of the recommendation model from – e.g., the reviews – and the aspects can be paired up with consumer opinions to express a clear sentiment on the aspect.

---

情感分析领域的研究人员探索了数据挖掘和机器学习技术来进行aspect-level 情感分析，目的是从文本中提取aspect-sentiment 对。

Researchers in the sentiment analysis community have explored both data mining and machine learning techniques for aspect-level sentiment analysis, which aims to extract aspect-sentiment pairs from text.

例如，Hu and Liu(2004)提出了一种频繁特征挖掘（frequent feature mining）方法来进行aspect-level 情感分析。 Lu et al. (2011)提出了一个自动构建上下文感知的情感词典的优化方法。 并且Liu（2012）情感分析和意见挖掘进行综述。

For example, Hu and Liu (2004) proposed a frequent feature mining approach to aspect-level sentiment analysis, and Lu et al. (2011) proposed an optimization approach to construct context-aware sentiment lexicons automatically.A comprehensive review on sentiment analysis and opinion mining is summarized in Liu (2012）

基于以上研究工作，Zhang et al.（2014b）(就是作者他自己)开发了parse-level 情感分析工具，叫做Sentires ，可以从产品领域的评论中抽取aspect–opinion–sentiment 三元组。例如，用户对手机的大规模评论，该工具包可以提取“noise–high–negative”, “screen–clear–positive”, and “battery_life–long–positive” 等三元组。

Based on these research efforts, Zhang et al. (2014b) developed a phrase-level sentiment analysis toolkit named Sentires1to extract “aspect–opinion–sentiment” triplets from reviews of a product domain. For example, given large-scale user reviews about mobile phones, the toolkit can extract triplets such as “noise–high–negative”, “screen clear–positive”, and “battery_life– long–positive”

工具包也可以直接检测不同aspect words 下的意见词（opinion words）的上下文情感（contextual sentiment ）。例如，‘noise’ 和 ‘high’ 成对 通常表示消极情感，但是’quality’和’high’ 成对表示积极情感。利用程序构建aspect–opinion–sentiment 三元组，以后可以在评论中检测其中包含哪一个三元组。

The toolkit can also detect the contextual sentiment of the opinion words under different aspect words. For example, though “noise” paired with “high” usually represents a negative sentiment, when “quality” is paired with “high”, however, it instead shows a positive sentiment. Based on the dictionary of aspect–opinion–sentiment triplets constructed by the program, it can further detect which triplets are contained in a review sentence.

---

基于这个工具包，研究者开发了不同的可解释推荐模型。
例如Zhang et al.（2014a）提出了explicit factor model , 使用aspect-opinion 词云作为解释，来突出推荐的item在某方面的表现，例如，“bathroom-clean”。

For example, Zhang et al. (2014a) proposed an explicit factor model for explainable recommendation, which presents aspect-opinion word clouds as explanations, such as “bathroom–clean”, to highlight the performance of the recommended item on certain aspects.

Wu and Ester (2015）在TripAdvisor的可解释酒店推荐上，开发了主题建模（topic modeling）方法，该方法对三个酒店特征（位置，服务，房间），生成主题词云的解释模型，如图2.6.词云中词的大小与该方面的情感观点成正比。

Wu and Ester (2015) developed a topic modeling approach for explainable hotel recommendations on TripAdvisor, which generates topical word cloud explanations on three hotel features (Location, Service, and Room), as shown in Figure 2.6. The word size in the word cloud is proportional to the sentiment opinion of the aspect.

Ren et al.（2017）提出了基于用户观点和社会关系的，social collaborative viewpoint regression（sCVR）model，来预测item 评分。这个模型以viewpoints 作为解释，而viewpoint 指的是具有特定情感标签的主题（a topic with a specific sentiment label）。结合了可信的用户关系，模型帮助用户理解他们朋友关于特定item的意见。

Ren et al. (2017) proposed a social collaborative viewpoint regression (sCVR) model for predicting item ratings based on user opinions and social relations, which provides viewpoints as explanations, where a viewpoint refers to a topic with a specific sentiment label. Combined with trusted user relations, it helps users to understand their friends’ opinion about a particular item.

Wang et al.(2018b) 开发了多任务学习的可解释推荐方案，通过联合张量分解框架，将用于推荐的用户偏好建模，和用于解释的内容建模学习任务集成在一起。更多细节在第三章。

Wang et al. (2018b) developed a multi-task learning solution for explainable recommendation, where two companion learning tasks of user preference modeling for recommendation and opinionated content modeling for explanation are integrated via a joint tensor factorization framework. We will introduce more about the model details in the explainable recommendation model section (Section 3).

2.4 Sentence Explanation (display style)

sentence-level 方法为用户提供解释性的句子。
该方法可进一步分为template-based 方法和generation-based方法。

Sentence-level approach provides explanation sentences to users. This approach can be further classified into template-based approach and generation-based approach.

1. template -based
   基于模板的方法首先定义解释句子的模板，然后用针对不同用户的个性化词汇填充模板。

   Template-based approach first defines some explanation sentence templates, and then fills the templates with different words to personalize them for different users.

   例子：

   基于句子模板，填充feature

   For example, Zhang et al. (2014a) constructed explanations by telling the user that You might be interested in feature, on which this product performs well. In this template, the feature will be selected based on personalization algorithms to construct a personalized explanation, as shown in Figure 2.7. Based on the templates, the model can also provide “dis-recommendations” to let the user know why an item is not a good fit, by telling the user You might be interested in feature, on which this product performs poorly. Based on user studies, it shows that providing both recommendation and dis-recommendation explanations improve the persuasiveness and conversion rate of recommender systems.
   

   基于句子模板，填充feature 和opinion

   Wang et al. (2018b) provided template-based explanations based on both feature and opinion words, for example, an explanation for Yelp restaurant recommendation could be Its decor is [neat] [good] [nice]. Its sandwich is [grilled] [cajun] [vegan]. Its sauce is [good] [green] [sweet], where words in brackets are opinion words selected by the model to describe the corresponding item feature.

   基于句子模板，填充feature和modifier

   Tao et al. (2019a) further integrated regression trees to guide the learning of latent factor models, and used the learnt tree structure to explain the recommendations. They added predefined modifiers in front of the selected features to construct template-based explanations, such as We recommend this item to you because its [good/excellent] [feature] matches with your [emphasize/taste] on [feature].

   基于句子模板，在知识图谱中，通过多视角学习层级feature。

   Gao et al. (2019) proposed an explainable deep multi-view learning framework to model multi-level features for explanation. They also adopted feature-based templates to provide explanations, and the features are organized in an industry-level hierarchy named Microsoft Concept Graph. The model improves accuracy and explainability simultaneously, and is capable of providing highly usable explanations in commercial systems. Technical details of the above models will be introduced in the explainable recommendation model section.

2. generation-based

   基于自然语言生成技术，可以不用模板直接生成句子。

   Based on natural language generation techniques, we can also generate explanation sentences directly without using templates.

   例子：

   For example, Costa et al. (2018) generated explanation sentences based on long-short term memory (LSTM). By training over large-scale user reviews, the model can generate reasonable review sentences as explanations, as shown in Figure 2.8.
   

   通过众包和计算生成个性化解释

   Inspired by how people explain word-of-mouth recommendations, Chang et al. (2016) proposed a process to combine crowdsourcing and computation to generate personalized natural language explanations. The authors also evaluated the generated explanations in terms of efficiency, effectiveness, trust, and satisfaction.

   通过GRU生成推荐提示

   Li et al. (2017) leveraged gated recurrent units (GRU) to generate tips for a recommended restaurant in Yelp. According to the predicted ratings, the model can control the sentiment attitude of the generated tips, which help users understand the key features of the recommended items.

   多任务推荐模型，通过sequence to sequence的生成器

   Lu et al. (2018b) proposed a multi-task recommendation model, which jointly learns to perform rating prediction and recommendation explanation. The explanation module employs an adversarial sequence to sequence learning technique to encode, generate, and discriminate the user and item reviews. Once trained, the generator can be used to generate explanation sentences.

---

但是又有问题来了，许多生成方法都是用用户评论作为语料训练，但是用户评价是noisy的，并不是所有句子都是用户对自己决策过程的解释。

A lot of explanation generation approaches rely on user reviews as the training corpus to train an explanation generation model. However, user reviews are noisy, because not all of the sentences in a review are explanations or justifications of the users’ decision-making process.

解决的例子：

Motivated by this problem, Chen et al. (2019a) 就是作者proposed a hierarchical sequence-to-sequence model (HSS) for personalized explanation generation, which includes an auto-denoising mechanism that selects sentences containing item features for model training.

Ni et al. (2019) introduced new datasets and methods to address this recommendation justification task. In terms of data, the authors proposed an extractive approach to identify review segments that justify users’ intentions. In terms of generation, the authors proposed a reference-based Seq2Seq model with aspect-planning to generate explanations covering different aspects. The authors also proposed an aspect-conditional masked language model to generate diverse justifications based on templates extracted from justification histories.

2.5 Visual Explanation（display style）

研究者利用视觉图像的直观性进行可解释性推荐。

To leverage the intuition of visual images, researchers have tried to utilize item images for explainable recommendation.

In Figure 2.1, for example, to explain to Bob that the lens is recommended because of its collar appearance, the system highlights the image region corresponding to the necklet of the lens.

---

Lin et al. (2019) studied the explainable outfit recommendation problem, for example, given a top, how to recommend a list of bottoms (e.g., trousers or skirts) that best match the top from a candidate collection, and meanwhile generate explanations for each recommendation.
Technically, this work proposed a convolutional neural network with a mutual attention mechanism to extract visual features of the outfits, and the visual features are fed into a neural prediction network to predict the rating scores for the recommendations.
During the prediction procedure, the attention mechanism will learn the importance of different image regions, which tell us which regions of the image are taking effect when generating the recommendations, as shown in Figure 2.9(a).

Chen et al. (2019b) proposed visually explainable recommendation based on personalized region-of-interest highlights, as shown in Figure 2.10. The basic idea is that different regions of the product image may attract different users.
As shown by the example in Figure 2.9(b), even for the same shirt, some users may care about the collar design while others may pay attention to the pocket.
As a result, the authors adopted a neural attention mechanism integrated with both image and review information to learn the importance of each region of an image.
The important image regions are highlighted in a personalized way as visual explanations for users.

---

总结，视觉上解释推荐还处于起步阶段。

In general, the research on visually explainable recommendation is still at its initial stage. With the continuous advancement of deep image processing techniques, we expect that images will be better integrated into recommender systems for both better performance and explainability.

2.6 Social Explanation（display style）

利用relevant-user进行解释，对用户来说有可信度和隐私问题。（2.1节已介绍）
但是如果我们告诉用户，他的朋友们对推荐的商品有相似的兴趣，那么这解释容易被接受。 所以提出利用社交信息的推荐。

As discussed in the previous subsections, a problem of relevant-user explanations is trustworthiness and privacy concerns, because the target user may have no idea about other users who have “similar interests”
Usually, it will be more acceptable if we tell the user that his/her friends have similar interests in the recommended item. As a result, researchers proposed to generate social explanations, that is, explanations with the help of social information

---

例子：

基于human-style、item-style， feature style 和 hybrid-style 在社交推荐系统中进行解释。

Papadimitriou et al. (2012) studied human-style, item-style, feature style, and hybrid-style explanations in social recommender systems; they also studied geo-social explanations to combine geographical data with social data.
For example, Facebook provides friends in common as explanations when recommending a new friend to a user (see Figure 2.11(b)).

该研究者在音乐推荐上展示有多少个朋友喜欢这个歌曲。这会影响用户点击查看率，而与评分无关。

Sharma and Cosley (2013) studied the effects of social explanations in music recommendation by providing the target user with the number of friends that liked the recommended item (see Figure 2.11(b)). The authors found that explanations influence the likelihood of users checking out the recommended artists, but there is little correlation between the likelihood and the actual rating for the artist.

Chaney et al. (2015) presented social Poisson factorization, a Bayesian model that incorporates a user’s preference with her friends’ latent influence, which provides explainable serendipity to users, i.e., pleasant surprise due to novelty.

---

Except for friend recommendation, social explanations also take effect in other social network scenarios.

For example, Park et al. (2018) proposed a unified graph structure to exploit both rating and social information to generate explainable product recommendations. In this framework, a recommendation can be explained based on the target user’s friends who have similar preferences, as shown in Figure 2.12.

Quijano-Sanchez et al. (2017) introduced a social explanation system applied to group recommendation, which significantly increased the likelihood of the user acceptance, the user satisfaction, and the system efficiency to help users make decisions.

Wang et al. (2014) generates social explanations such as “A and B also like the item” . They proposed to find an optimal set of users as the most persuasive social explanation. Specifically, a two-phase ranking algorithm is proposed, where the first phase predicts the persuasiveness score of a single candidate user, and the second phrase predicts the persuasiveness score of a set of users based on the predicted persuasiveness of the individual users, by taking the marginal net utility of persuasiveness, credibility of the explanation and reading cost into consideration.

2.7 Summary

1. Explanations based on relevant users or item
   用最近邻的users或items作为解释。
   常用协同过滤的方法

   which present nearest-neighbor users or items as an explanation.
   They are closely related to the critical idea behind user-based or item-based collaborative filtering methods.

2. Feature-based explanation
   将item特征与用户兴趣画像进行匹配
   常用基于内容的推荐方法

   which provides users with the item features that match the target user’s interest profile. This approach is closely related to content-based recommendation methods.

3. Opinion-based explanation
   从用户生成的文本中，以用户的集体意见作为解释

   which aggregates users’ collective opinions in user generated contents as explanations.

4. Textual sentence explanation
   display style：呈现解释性的句子。
   句子由模板产生或者由语言生存模型生成

   which provides the target users with explanation sentences.
   The sentence could be constructed based on pre-defined templates or directly generated based on natural language generation models.

5. Visual explanations,
   呈现基于图像的解释。
   解释可以是整个图像或者图像的一部分。

   which provide users with image-based explanations.
   The visual explanation could be a whole image or a region-of-interest highlight in the image.

6. Social explanations,
   由用户社交关系提供解释。

   which provide explanations based on the target user’s social relations.
   They help to improve user trust in recommendations and explanations.

前三个和推荐算法有关，后三个解释呈现给用户的方式有关，而不拘泥于某一算法。

It should be noted that while 1 to 3 are usually bonded with certain types of recommendation algorithms, 4 to 6 focus more on how the explanations are shown to users, which are not necessarily generated by one particular type of algorithm.