2019_WSDM_Social Attentional Memory Network Modeling Aspect- and Friend-level Differences in Recomme

[论文阅读笔记]2019_WSDM_Social Attentional Memory Network Modeling Aspect- and Friend-level Differences in Recommendation

论文下载地址： https://doi.org/10.1145/3289600.3290982
发表期刊：WSDM
Publish time: 2019
作者及单位:

• Chong Chen DCST, Tsinghua University Beijing, China cc17@mails.tsinghua.edu.cn
• Min Zhang∗ DCST, Tsinghua University Beijing, China z-m@tsinghua.edu.cn
• Yiqun Liu DCST, Tsinghua University Beijing, China yiqunliu@tsinghua.edu.cn
• Shaoping Ma DCST, Tsinghua University Beijing, China msp@tsinghua.edu.cn

数据集： 正文中的介绍

• Delicious https://grouplens.org/datasets/hetrec-2011/
• Ciao http://www.jiliang.xyz/trust.html
• Epinions https://alchemy.cs.washington.edu/data/epinions/

代码：

• （文中作者给的）

其他：

其他人写的文章

• 《Social Attentional Memory Network: Modeling Aspect- andFriend-level Differences in Recommendation》
• 论文笔记（Social Attentional Memory Network：Modeling Aspect- and Friend-level Differences in Recom-)
• 推荐系统之社会化推荐经典论文总结

简要概括创新点： aspect-level差异用attention-based memory module来解决，friend-level就是个注意力机制。 其实主要还是想引入Memory Network。之所以用MF，BPR这种最原始的模型（单比是不如GNN， HGNN…的）是为了迁就Memory Network， user-item 的rating都只有0和1（买或没买，click或unclick）

• we propose to model both aspect-level differences and friend-level differences for improving the performance of social-aware recommendation. (为了提高社交感知推荐的性能，我们提出对方面级差异和朋友级差异进行建模。)
• In this paper, we present a Social Attentional Memory Network (or SAMN for short), which utilizes the recent advances in memory networks [23,31] and neural attention mechanisms [2, 3, 34]. (在本文中，我们提出了一个社会注意记忆网络（简称SAMN），它利用了记忆网络[23,31]和神经注意机制[2,3,34]的最新进展。)
  • Specifically, we first design an attention-based memory module to learn the user-friend specific relation vectors, and then employ friend-level attention to automatically select informative friends for user preference modeling. (具体来说，我们首先设计了一个基于注意的记忆模块来学习用户朋友特定的关系向量，然后利用朋友级别的注意自动选择信息性朋友进行用户偏好建模。)
  • The memory component allows reading and writing operations to encode complex user and friend relations. (记忆组件允许读写操作来编码复杂的用户和朋友关系。)
  • An associative attention-based addressing scheme places higher weights on aspects in which user and his friend share similar preferences. (基于关联注意的寻址方案在用户和朋友有相似偏好的方面赋予更高的权重。)
  • The attention-based memory module is controlled by the user-friend interaction, making the learned relation vector corresponds to each user-friend pair. (基于注意的记忆模块由用户-朋友交互控制，使学习到的关系向量对应于 每个用户-朋友对。)
  • In the friend-level attention modeling process, a two-layer attention network is adopted to model the influence strength among users’ friends in a distant supervised manner. (在朋友级注意建模过程中，采用 两层注意网络，以 远程监督的方式 对用户朋友之间的影响强度进行建模。)
  • Then, the two components are fused together to mutually enhance each other via an end-to-end training process. (然后，这两个组件通过端到端的培训过程融合在一起，相互增强。)
• 细节
• In our work, we also adopt BPR as our basic learning model because of its effectiveness in exploiting the unobserved user-item feedback. (在我们的工作中，我们还采用 BPR 作为我们的基本学习模型，因为它可以有效地利用 未观察到的用户项目反馈。)
• All the datasets were preprocessed to make sure that all items have at least five ratings. The statistical details of these datasets are presented in Table 1. (所有数据集都经过预处理，以确保所有项目至少有五个评分。这些数据集的统计细节见表1。)
• Since we focus on the implicit feedback, we transform the detailed ratings into a value of 0 or 1 indicating whether the user has rated the item. (该数据集包含用户对他们购买的物品的评分以及用户之间的社会关系。由于我们关注的是隐性反馈，因此我们将详细评分转换为0或1的值，以指示用户是否对该项目进行了评分。)
• The model is optimized by pairwise ranking and negative sampling strategy (通过成对排序和负采样策略对模型进行优化)
• To optimize the objective function, we adopt mini-batch Adagrad [7] as the optimizer.
• Our prediction part is built on Matrix Factorization (MF)

ABSTRACT

• (1) Social connections are known to be helpful for modeling users’ potential preferences and improving the performance of recommender systems. (众所周知，社交关系有助于模拟用户的潜在偏好，提高推荐系统的性能。)
• (2) However, in social-aware recommendations, there are two issues which influence the inference of users’ preferences, and haven’t been well-studied in most existing methods: (然而，在具有社会意识的推荐中，有两个问题会影响用户偏好的推断，并且在大多数现有方法中没有得到很好的研究：)
  • First, the preferences of a user may only partially match that of his friends in certain aspects, especially when considering a user with diverse interests. (首先，用户的偏好可能只在某些方面与朋友的偏好部分匹配，尤其是在考虑具有不同兴趣的用户时。)
  • Second, for an individual, the influence strength of his friends might be different, as not all friends are equally helpful for modeling his preferences in the system. (第二，对于一个人来说，朋友的影响力可能不同，因为并非所有的朋友都能同样有效地在系统中为他的偏好建模。)
• (3) To address the above issues, in this paper, we propose a novel Social Attentional Memory Network (SAMN) for social-aware recommendation. (为了解决上述问题，在本文中，我们提出了一种新的社会注意记忆网络（SAMN），用于社会感知推荐。)
  • Specifically, we first design an attention-based memory module to learn user-friend relation vectors, which can capture the varying aspect attentions that a user share with his different friends. (具体来说，我们首先设计了一个 基于注意的记忆模块 来学习用户-朋友关系向量，它可以捕捉用户与不同朋友分享的 不同方面的注意。)
  • Then we build a friend-level attention component to adaptively select informative friends for user modeling. (然后，我们构建了一个 朋友级注意组件 ，自适应地选择信息丰富的朋友 进行用户建模。)
  • The two components are fused together to mutually enhance each other and lead to a finer extended model. (这两个组件融合在一起，相互增强，形成更精细的扩展模型。)
• (4) Experimental results on three publicly available datasets show that the proposed SAMN model consistently and significantly outper- forms the state-of-the-art recommendation methods. Furthermore, qualitative studies have been made to explore what the proposed attention-based memory module and friend-level attention have learnt, which provide insights into the model’s learning process. (在三个公开数据集上的实验结果表明，所提出的SAMN模型一致且显著地超过了最先进的推荐方法。此外，还进行了定性研究，以探索所提出的基于注意的记忆模块和朋友级注意学习了什么，从而深入了解了该模型的学习过程。)

CCS CONCEPTS

• Information systems → Recommender systems; • Computing methodologies → Neural networks;

KEYWORDS

Recommender Systems, Collaborative Filtering, Attention, Memory Networks, Social Connections

1 INTRODUCTION

• (1) Users expect personalized products and information in modern E-commerce, entertainment and social media platforms. In this case, recommender systems are designed to generate personalized item recommendations and deal with the information overload problem. Many recommendation methods are based on Collaborative Filtering (CF) [14, 17, 18, 26], which mainly makes use of users’ historical records such as ratings, clicks, and purchases. (用户希望在现代电子商务、娱乐和社交媒体平台上获得个性化的产品和信息。在这种情况下，推荐系统旨在生成个性化的商品推荐，并处理信息过载问题。许多推荐方法都基于 协同过滤（CF） [14,17,18,26]，它主要利用用户的历史记录，如评分、点击和购买。)

• (2) Recently, owing to the prevalence of social media, many E-commerce sites have become popular social platforms that help users discuss and select items [25], such as Delicious, Ciao and Epinions. (最近，由于社交媒体的普及，许多电子商务网站已经成为流行的社交平台，帮助用户讨论和选择项目[25]，如Delicious、Ciao和ePionions。)

  • In these social applications, users like to spread their preferences of items to their social connections, and a user’s preferences can not only be inferred from the items he bought and clicked, but also can be inferred from his social connections. (在这些社交应用程序中，用户喜欢将他们对项目的偏好传播到他们的社交关系中，用户的偏好不仅可以从他购买和单击的项目中推断出来，还可以从他的社交关系中推断出来。)
  • Generally, there are two types of social connections: (一般来说，有两种类型的社会关系：)
    • friends in undirected social networks (无定向社交网络中的朋友)
    • and followers in directed social networks. (以及定向社交网络中的追随者。)
  • Since we do not focus on the differences between the two types, for the convenience of description, we refer to the user’s social connections as friends in this paper. As shown in previous studies on social-aware recommender systems, the social behavior of users and their interactions with items are positively correlated [22, 25, 37, 42]. By considering users’ social connections, social-aware methods can utilize a much larger volume of data to tackle the data sparsity issue, and further improve the performance of recommender systems. (由于我们不关注这两种类型之间的差异，为了便于描述，我们在本文中将用户的社交关系称为朋友。正如之前关于社交感知推荐系统的研究所显示的，用户的社交行为及其与物品的互动 正相关 [22,25,37,42]。通过考虑用户的社交关系，社交感知方法可以利用更大的数据量来解决数据稀疏性问题，并进一步提高推荐系统的性能。)

• (3) However, in social-aware recommendations, there are two issues which influence the inference of users’ preferences, and haven’t been well-studied in most existing methods. (然而，在社交感知推荐中，有两个问题会影响用户偏好的推断，并且在大多数现有方法中都没有得到很好的研究。)

• (4) The first one is aspect-level differences. Generally, users and their friends only have the same preferences in certain aspects. It is well recognized that the preference of a user can be used to infer his friends’ preference and vice versa, which could be denoted as an influence vector. (第一个是方面层面的差异。一般来说，用户和他们的朋友只在某些方面有相同的偏好。众所周知，用户的偏好可以用来推断朋友的偏好，反之亦然，这可以表示为影响向量。)

  • Existing methods like [19, 29, 38] assume that this vector keeps the same when facing different friends. However, a user may pay the most attention to one aspect for a friend but focus on another aspect for a different friend. (像[19,29,38]这样的现有方法假设，当面对不同的朋友时，这个向量保持不变。然而，对于一个朋友，用户可能最关注一个方面，但对于另一个朋友，用户可能会关注另一个方面。)
  • In Figure 1 we show an example that is common in social relationships. User A is a friend of user B and user C, but the reasons are distinct: user A and B are friends because they are both interested in movie, while A and C are friends because they both like basketball. (在图1中，我们展示了一个社会关系中常见的例子。用户A是用户B和用户C的朋友，但原因是截然不同的：用户A和B是朋友，因为他们都对电影感兴趣，而A和C是朋友，因为他们都喜欢篮球。)
  • The aspect-level differences should be considered when building a social-aware recommender system. However, since users’ preference inference is usually complex and non-linear, the aspect differences are hard to be captured by traditional latent factor based models. (在构建具有社会意识的推荐系统时，应考虑方面层面的差异。 然而，由于用户的偏好推理通常是复杂和非线性的，传统的基于潜在因素的模型很难捕捉到方面的差异。)

• (5) The second issue is friend-level differences. For a user, the influence strength of his friends should be different and dynamic. Each user is associated with a set of friends in social networks, but it does not necessarily indicate that every friend has equal influence strength on his behaviors. (第二个问题是朋友级别的差异。对于用户来说，朋友的影响力应该是不同的、动态的。在社交网络中，每个用户都与一组朋友联系在一起，但这并不一定表明每个朋友对自己的行为都有同等的影响力。)

  • For example, when buying basketball shoes, a user will follow the advice of his friends who play basketball, but when it comes to a trip, he will turn to those who love traveling. (例如，在购买篮球鞋时，用户会听从打篮球的朋友的建议，但在旅行时，他会转向那些喜欢旅行的人。)
  • In previous work, the social influence strength is usually set equally for the social connections [15], or relied on a predefined static function [9, 16]. These settings are not robust in real life. To better characterize a user’s preferences, the model requires different attentions on the set of the user’s friends. (在以前的工作中，社会影响强度通常是为社会关系 同等 设置的[15]，或者依赖于预定义的 静态 函数[9,16]。 这些设置在现实生活中并不可靠。为了更好地描述用户的偏好，该模型需要对用户的朋友集进行不同的关注。)

• (6) Motivated by the above observations,

• we propose to model both aspect-level differences and friend-level differences for improving the performance of social-aware recommendation. (为了提高社交感知推荐的性能，我们提出对方面级差异和朋友级差异进行建模。)

• In this paper, we present a Social Attentional Memory Network (or SAMN for short), which utilizes the recent advances in memory networks [23,31] and neural attention mechanisms [2, 3, 34]. (在本文中，我们提出了一个社会注意记忆网络（简称SAMN），它利用了记忆网络[23,31]和神经注意机制[2,3,34]的最新进展。)

  • Specifically, we first design an attention-based memory module to learn the user-friend specific relation vectors, and then employ friend-level attention to automatically select informative friends for user preference modeling. (具体来说，我们首先设计了一个基于注意的记忆模块来学习用户朋友特定的关系向量，然后利用朋友级别的注意自动选择信息性朋友进行用户偏好建模。)
  • The memory component allows reading and writing operations to encode complex user and friend relations. (记忆组件允许读写操作来编码复杂的用户和朋友关系。)
  • An associative attention-based addressing scheme places higher weights on aspects in which user and his friend share similar preferences. (基于关联注意的寻址方案在用户和朋友有相似偏好的方面赋予更高的权重。)
  • The attention-based memory module is controlled by the user-friend interaction, making the learned relation vector corresponds to each user-friend pair. (基于注意的记忆模块由用户-朋友交互控制，使学习到的关系向量对应于 每个用户-朋友对。)
  • In the friend-level attention modeling process, a two-layer attention network is adopted to model the influence strength among users’ friends in a distant supervised manner. (在朋友级注意建模过程中，采用 两层注意网络，以 远程监督的方式 对用户朋友之间的影响强度进行建模。)
  • Then, the two components are fused together to mutually enhance each other via an end-to-end training process. (然后，这两个组件通过端到端的培训过程融合在一起，相互增强。)

• We evaluate SAMN extensively on three real-world datasets. Experimental results show that our model consistently outperforms the state-of-the-art methods, and also verify the effectiveness of our designed attention component and memory network. (我们在三个真实数据集上对SAMN进行了广泛的评估。实验结果表明，我们的模型始终优于最先进的方法，也验证了我们设计的注意成分和记忆网络的有效性。)

• (7) The main contributions of this work are summarized as follows. (这项工作的主要贡献总结如下。)

  • (1) We propose a new model for social-aware recommender systems, which considers both aspect-level differences among user-friend co-preferences and friend-level differences on social influence strength. (我们提出了一个新的社会感知推荐系统模型，该模型既考虑了用户朋友共同偏好的方面水平差异，也考虑了朋友水平差异对社会影响力的影响。)
  • (2) To the best of our knowledge, we are the first to employ an attention-based memory module to construct user-friend specific relation vector. (据我们所知，我们是第一个使用基于注意的记忆模块来构建用户朋友特定关系向量的人)
    • We also introduce the friend-level attention to adaptively measure the social influence strength among users’ friends. (我们还引入了朋友级注意，以自适应地测量用户朋友之间的社会影响力强度)
    • These two parts are fused in a unified framework and can be learned through efficient end-to-end training. (这两部分融合在一个统一的框架中，可以通过有效的端到端训练来学习。)
  • (3) Through extensive experiments conducted on three benchmark datasets, we show that SAMN consistently outperforms the state-of-the-art models. (通过在三个基准数据集上进行的大量实验，我们表明SAMN始终优于最先进的模型。)

2 RELATED WORK

2.1 Traditional Collaborative Filtering

• (1) Among the various collaborative filtering methods, matrix factorization (MF) is the most popular one, and is also the basis of many effective recommender models [27, 30]. Popularized by the Netflix Challenge, early MF methods [18] were designed to model users’ explicit feedback by mapping users and items to a latent factor space, such that user-item relationships (ratings) can be obtained by their latent factors’ dot product. (在各种 协同过滤方法 中，矩阵分解（MF） 是最流行的一种，也是许多有效推荐模型的基础[27,30]。通过Netflix挑战推广，早期的MF方法[18]被设计为通过将用户和项目映射到潜在因素空间来模拟用户的显式反馈，这样用户-项目关系（评级）可以通过其潜在因素的点积获得。)

• (2) Later on, some researchers found that a well-designed MF model in rating prediction may not perform well in Top-K recommendation , and called on recommendation research to focus more on the ranking task [6]. (后来，一些研究人员发现，评级预测中设计良好的MF模型在 Top-K推荐 中可能表现不佳，并呼吁推荐研究更多地关注 排名任务[6]。)

  • In this case, Rendle et al. [26] first proposed a pair-wise learning method BPR, which is a sample-based method that optimizes the model based on the relative preference of a user over pairs of items. (在这种情况下，Rendle等人[26]首先提出了一种 成对学习方法BPR，这是一种 基于样本 的方法，根据用户相对于 成对项目 的相对偏好来优化模型。)
  • Then, the pairwise learning strategy has been widely used to optimize recommender models [3, 34, 39–41]. and become a dominant technique in recommendation. (然后，成对学习策略 被广泛用于优化推荐模型[3,34,39–41]。并成为推荐中的主导技术。)

• In our work, we also adopt BPR as our basic learning model because of its effectiveness in exploiting the unobserved user-item feedback. (在我们的工作中，我们还采用 BPR 作为我们的基本学习模型，因为它可以有效地利用 未观察到的用户项目反馈。)

2.2 Social-aware Recommendation

• (1) In the last few years, there is a large literature exploiting users’ social connections for improving the recommendation performance [25]. Most studies assumed that a user’s decision can be affected by his friends’ opinions and behaviors. E.g., (在过去几年中，有大量文献利用用户的社会关系来提高推荐性能[25]。大多数研究都假设用户的决定会受到朋友的意见和行为的影响。)

  • in [42] , the authors assumed that users are more likely to have seen items consumed by their friends, and exploited this effect to extend BPR [26] by changing the negative sampling strategy. (在[42]中，作者假设用户更可能看到他们的朋友消费的物品，并利用这一效应通过改变 负采样策略 来扩展 BPR [26]。)
  • Jamali et al. designed a social influence propagation based model in latent based recommendation models (SocialMF) [15]. (Jamali等人在 基于潜在推荐模型（SocialMF）中设计了一个 基于社会影响传播的模型 [15]。)
  • Based on a generative influence model, the work [38] exploits social influence from friends for item recommendation by leveraging information embedded in the user social network. (基于 生成性影响模型 ，这项工作[38]利用嵌入用户社交网络中的信息，利用朋友的社会影响进行商品推荐。)
  • However, many existing methods [19, 29, 38] assume that users’ influence vectors stay the same for different friends, thus the aspect-level differences are not well-studied. (然而，许多现有的方法[19,29,38]假设用户对不同朋友的影响向量保持不变，因此没有很好地研究 层面的差异。)

• (2) In social-aware recommendation, social influence strength modeling is a central problem [8, 9]. E.g., (在社会意识推荐中，社会影响强度 建模是一个核心问题[8,9]。)

  • Goyal et al. [9] designed a model to calculate influence strength from users’ historical behaviors. (Goyal等人[9]设计了一个模型，根据用户的历史行为计算影响力。)
  • However, in most of the existing methods, the social influence strength is assumed equal among friends [15] or with a simple metric from other sources [9, 16] (e.g., the strength between their interactions in the past). (然而，在大多数现有方法中，假设朋友之间的 社会影响强度相等[15]，或者使用来自其他来源的简单度量[9,16]（例如，他们过去互动之间的强度）。)

• (3) To the best of our knowledge, few has explored the neural networks for modeling aspect-level differences and friend-level differences in social-aware recommendation, which is the main concerns of our work. (就我们所知，很少有人探索神经网络来模拟社交感知推荐中的方面水平差异和朋友水平差异，这是我们工作的主要关注点。)

2.3 Deep Learning in Recommendation

• (1) Recently, deep learning has yielded an immense success in many fields like computer vision, speech recognition and natural language processing [20]. (最近，深度学习在计算机视觉、语音识别和自然语言处理等领域取得了巨大成功[20]。)

• Some researchers also tried to exploit different neural network structure for improving the performance of recommendations. (一些研究人员还试图利用不同的神经网络结构来提高推荐的性能。)

  • In [11]. He et al. presented a **Neural Collaborative Filtering (NCF) framework to address implicit feedback by jointly learning a matrix factorization and a feedforward neural network, NCF is also the state-of-art recommendation method for using only user-item historical records. (在[11]中。他等人提出了一个 神经协同过滤(NCF) 框架，通过联合学习 矩阵分解 和 前馈神经网络 来解决隐式反馈，NCF也是仅使用用户项历史记录的最先进推荐方法。)
  • Later, Neural Factorization Machines (NFM) [10] was developed to enhance FM by 0modeling higher-order and non-linear feature interactions. (后来，开发了 神经因子分解机（NFM） [10]，通过对 高阶 和 非线性 特征交互进行建模来增强FM。)
  • More recently, [32] presents an attentive recurrent network for temporal social-aware recommendation (ARSE). There are two major differences between our work and ARSE: (最近，[32]提出了一个 时态社会意识推荐(ARSE) 注意力循环网络。我们的工作和ARSE之间有两个主要区别：)
    • (1) We focus on a more general problem while ARSE focuses on temporal recommendations via Recurrent Neural Network (RNN) and attention mechanisms. (我们关注的是一个更普遍的问题，而ARSE则关注通过 循环神经网络（RNN） 和 注意机制的时间建议。)
    • (2) Our work introduces memory network to address the problem of aspect-level differences between users and their friends. (我们的工作引入了 记忆网络 来解决用户和他们的朋友之间的方面级别差异问题。)

• (2) Attention mechanism has been shown effective in many machine learning tasks such as image captioning and machine translation [1, 28]. (注意机制已被证明在许多机器学习任务中有效，如图像字幕和机器翻译[1,28]。)

  • In the field of recommendation, [3] introduced both component-level and item-level attention into a CF framework for multimedia recommendation. (在推荐领域，[3]在多媒体推荐的CF框架中引入了组件级和项目级注意。)
  • [36] improved FM by learning the importance of different feature interactions via a neural attention network. ([36]通过神经注意网络学习不同特征交互的重要性，改进了FM。)
  • Recently, Chen et al. [2] proposed to learn the “usefulness” of reviews with the help of attention mechanism for improving the performance and explainability of the recommender system. (最近，Chen等人[2]提出借助注意机制来学习评论的“有用性”，以提高推荐系统的性能和可解释性。)

• (3) Memory networks are recently introduced frameworks that combine reasoning, attention and memory for solving tasks in the areas of language understanding and dialogue. (记忆网络是最近引入的一种框架，将 推理、注意力 和 记忆 结合起来，用于解决 语言理解和对话 领域的任务。)

• It generally consist of two components: (它通常由两部分组成：)

  • an external memory typically a matrix (一种外部存储器，通常是一个矩阵)
  • and a controller which perform operations on the memory (e.g., read, write). (以及对存储器执行操作（例如，读、写）的控制器。)
  • The memory component increases model capacity independent of the controller while providing an internal representation of knowledge to track long-term dependencies and perform reasoning. (内存组件 独立于控制器增加模型容量，同时提供知识的内部表示，以 跟踪长期依赖关系 并 执行推理。)
  • The controller usually manipulates these memories with content-based addressing, which finds a scoring function between the given query and a passage of text [23, 31, 35]. (控制器 通常通过 基于内容的寻址 操作这些存储器，在给定的查询和一段文本之间找到一个评分函数[23,31,35]。)
  • For recommendations, [34] utilized memory module to learn the relationships between user-item interactions for extending Collaborative Metric Learning [13]. (对于推荐，[34]利用内存模块学习用户项交互之间的关系，以扩展 协同度量学习[13]。)
  • In [35], the authors proposed Collaborative Memory Network, while the associative addressing scheme of the memory module acts as a nearest neighborhood model identifying similar users. (在[35]中，作者提出了 协同记忆网络，而存储模块的关联寻址方案充当 识别相似用户的最近邻模型。)

3 SOCIAL ATTENTIONAL MEMORY

NETWORK (SAMN) In this section, we introduce our Social Attentional Memory Network (SAMN). （网络（SAMN）在本节中，我们将介绍我们的社会注意记忆网络（SAMN）。）

• First, we will present the general architecture of SAMN. (首先，我们将介绍SAMN的总体架构。)
• Then, we will show the detailed formulations of our proposed attention-based memory module and friend-level attention respectively, which are the main concerns in this paper. (然后，我们将分别展示我们提出的基于注意的记忆模块和朋友级注意的详细公式，这是本文的主要关注点。)
• Lastly we will go through the optimization details of SAMN. (最后，我们将介绍SAMN的优化细节)
  

3.1 Overview of SAMN

• The goal of our model is to make recommendations based on implicit feedback and social networks. (我们模型的目标是基于内隐反馈和社交网络提出建议。)
• Both aspect-level and friend-level differences are considered for improving the model performance and generalization. The architecture of the proposed model is shown in Figure 2. (为了提高模型的性能和泛化能力，同时考虑了方面级和朋友级的差异。拟议模型的架构如图2所示。)
• From the figure, we can make a simple high-level overview of our model: (从图中，我们可以对我们的模型进行简单的高层概述：)
  • (1) <font color=reUsers and items are converted to dense vector representations using an Embedding Layer. $u$ and $v$ are the user and item vectors respectively. (用户和项目使用嵌入层转换为密集向量表示。$u$和$v$分别是用户和项目向量。)
  • (2) The model contains two major components, which are (该模型包含两个主要部分)
    • the attention-based memory module (基于注意力的记忆模块)
    • and the friend-level attention component. (朋友级注意力组件。)
    • The attention-based memory module is designed for addressing the problem caused by aspect-level differences. （基于注意的记忆模块旨在解决因方面级别差异引起的问题。）
    • The friend vector $f_{(i,l)}$ is generated using a neural attention mechanism over an augmented memory matrix M. It dependents on user and friend, and is learned to represent the preference relationship between user and his friend. (友元向量$f_{(i,l)}$在 增强记忆矩阵$M$上使用神经注意机制生成。它依赖于用户和朋友，并学会表示用户和朋友之间的偏好关系。)
    • The friend-level attention is used to select informative friends for better inferring user preference. (朋友级注意用于选择信息丰富的朋友，以便更好地推断用户偏好。)
  • (3) The model is optimized by pairwise ranking and negative sampling strategy (通过成对排序和负采样策略对模型进行优化)

3.2 Attention-based Memory Module

• In most cases, users and their friends only have the same preferences in certain aspects, especially when considering a user with diverse interests. However, explicit relations between user-friend pairs are not available in implicit data (we don’t know what their shared interest aspects are). (在大多数情况下，用户和他们的朋友只在某些方面有相同的偏好，尤其是在考虑具有不同兴趣的用户时。然而，用户朋友对之间的显式关系在隐式数据中不可用（我们不知道他们的共同兴趣方面是什么）。)
• Motivated by the recent advance in memory network and attention mechanisms, we designed a new attention-based memory module to learn the relation vectors between users and their friends. (基于记忆网络和注意机制的最新发展，我们设计了一个新的基于注意的记忆模块来学习用户和朋友之间的关系向量。)
• The structure of the module is shown in Figure 3. The memory matrix of the module is represented as $M\in R^{N\times d}$, where $d$ is the dimension of the user and item embeddings and $N$ is the memory size. (模块的结构如图3所示。模块的内存矩阵表示为$M\in R^{N\times d}$，其中$d$是用户和项目嵌入的维度，$N$是记忆大小。)
• In matrix $M$, each slice is noted $M_j\in R^d$ as a memory slice. The input of the module is a user-friend pair $(u_i,u_{(i,l)})$, where $u_i$ denotes user $i$, and $u_{(i,l)}$ denotes the $l$-th friend of user $i$.
• The module returns the vector $f(i,l)$, which represents the relationship between $u_i$ and $u_{(i,l)}$. （模块返回向量 $f(i,l)$，它表示$u_i$ 和 $u_{(i,l)}$之间的关系。）

3.2.1 Joint Embedding.

• (1) Given the user-friend pair $(u_i,u_{(i,l)})$, the module first apply the following operation to learn a joint embedding of users and their friends. The denominator added in Eq. (1) is used to normalize and make the generated vectors have the same scale. (给定用户朋友对$(u_i,u_{(i,l)})$，模块首先应用以下操作来学习用户及其朋友的联合嵌入。公式(1)中添加的分母用于规范化并使生成的向量具有相同的比例)
  
  • where $\odot$ denotes the element-wise product of vectors. (其中$\odot$表示向量的元素乘积。)
  • The generated vector $s\in R^d$ is of the same dimension of $u_i$ and $u_{(i,l)}$. (生成的向量$s\in R^d$的维数与$u_i$ 和 $u_{(i,l)}$相同)
  • Note that other functions like the multi-layered perceptron (MLP) or just element-wise product without normalization can also be adopted, but we found that the method of Eq. (1) performs better. （请注意，也可以采用其他功能，如多层感知器（MLP）或无规范化的元素乘积，但我们发现等式（1）的方法性能更好。）

3.2.2 Key Addressing.

• (1) After we obtain the joint embedding vector $s$, the attention vector is learned from a key matrix $K\in R^{N\times d}$. Each element of the attention vector $\alpha$ is defined as: (在我们获得联合嵌入向量$s$后，注意向量密钥矩阵$K\in R^{N\times d}$学习。注意向量$\alpha$的每个元素定义如下：)
  
  • where $K_i\in R^d$ and the generated vector $\alpha \in R^N$.
• (2) Then the final attention scores are obtained by normalizing $\alpha$ using the softmax function: ($K_i\in R^d$和生成向量$\alpha \in R^N$。然后，通过使用softmax函数对$\alpha$进行归一化，得到最终的注意力分数：)
  

3.2.3 Generating Friend Vector.

• (1) In this step, the friend embedding $u_{(i,l)}$ is first extended to a matrix via the memory matrix $M$: (在这一步中，嵌入$u_{(i,l)}$的朋友首先通过记忆矩阵$M$扩展到一个矩阵：)
  

  • where $\odot$ denotes the element-wise product of vectors. (表示向量的元素乘积。)
  • The matrix $F\in R^{N\times d}$ can be interpreted as a storage of conceptual building blocks that used to describe the friend preferences in different latent aspects ($N$ can be seen as the number of latent aspects). (矩阵$F\in R^{N\times d}$可以解释为概念构建块的存储，用于描述不同潜在方面的朋友偏好（$N$可以被视为潜在方面的数量）。)
    

• (2) Finally, to generate the friend vector, we use the attention scores to calculate a weighted representation of $F$: (最后，为了生成朋友向量，我们使用注意分数来计算$F$的加权表示：)
  

• (3) The output is a specific relation vector $f_{(i,l)}$, which can be seen as the influence vector of user $i$’s $l$-th friend to user $i$’s preferences. (输出是一个特定的关系向量 $f_{(i,l)}$，它可以被视为用户$i$的第$l$个朋友对用户$i$偏好的影响向量。)

• (4) Let $f_{(i,1)},f_{(i,2)},...f_{(i,n)}$ be the the relation vectors of user $i$’s friends generated by the attentional memory module. As we have mentioned in section 1, generally not every friend has the same importance for inferring a user’s preference in real life. To address this problem, we introduce friend-level attention into our model, which can help to adaptively learn the influence strength of each friend. (让$f_{(i,1)},f_{(i,2)},...f_{(i,n)}$是由注意力记忆模块生成的用户$i$的朋友的关系向量。正如我们在第1节中提到的，通常不是每个朋友在推断用户在现实生活中的偏好时都具有相同的重要性为了解决这个问题，我们在我们的模型中引入了朋友级注意，这有助于自适应地学习每个朋友的影响力。)

3.3 Friend-level Attention

• (1) Attention mechanism has been widely adopted in many fields, such as computer vision [4], machine translation [1] and recommendation [2, 3, 36]. (注意机制已被广泛应用于许多领域，如计算机视觉[4]、机器翻译[1]和推荐[2,3,36]。)

• The goal of the friend-level attention is to assign non-uniform weights to users’ friends, and the weights are varied when the user interacts with different items. (朋友级注意的目标是为用户的朋友 分配非均匀的权重，当用户与不同的项目交互时，权重是不同的。)

• Intuitively, if a friend has more expertise on an item (or items of the similar type), he should have a larger influence on the user’s choice on the item. (直觉上，如果朋友对某个项目（或类似类型的项目）有更多的专业知识，他应该对用户对该项目的选择有更大的影响。)

• Formally, a two-layer network is applied to compute the attention score ${\beta }_{(i,l)}$ with user ($u_i$), current item ($v_j$) and friend vector ($f_{(i,l)}$) as inputs: (形式上，应用两层网络来计算注意力得分${\beta }_{(i,l)}$与用户 ($u_i$), 当前项目 ($v_j$) 和友元向量($f_{(i,l)}$)作为输入)

  • where $W_1\in R^{d\times k}$, $W_2\in R^{d\times k}$, $W_3\in R^{d\times k}b\in R^k$, $h\in R^k$ are model parameters,
  • $k$ denotes the dimension of attention network, (k表示注意网络的维度，)
  • and ReLU [24] is a nonlinear activation function.

• (2) Then, the final friend-level attention is normalized with a softmax function, which is a common practice in neural attention network. It makes the attention network a probabilistic interpretation, which can also deal with the problem that users may have different number of friends: (然后，使用softmax函数对最终的朋友级注意进行归一化，这是神经注意网络中的常见做法。它使注意力网络成为一种概率解释，也可以解决用户可能拥有不同数量朋友的问题：)
  

  • where $S_i$ denotes all friends that user $i$ has in the social network (其中，$S_i$表示用户$i$在社交网络中拥有的所有朋友)

• (3) After we obtain the attention weight of each friend, the final representation of user $i$ is through the sum: (在我们获得每个朋友的注意力权重后，用户$i$的最终表示形式是：)
  

  • which considers both the user’s own preference and the influence of his friends. ( 它同时考虑了用户自己的偏好和朋友的影响。)

• (4) Note that there have been a lot of work exploring the strategies of combine different features, such as concatenation, addition, or element-wise product. (请注意，已经有很多工作在探索组合不同功能的策略，例如 拼接 、添加 或 元素对应相乘 。)

• In this work, we adopt the addition fusion method, which has been applied in RBLT [33], NARRE [2] and A3NCF [5] and achieves good performance. (在这项工作中，我们采用了加法融合方法，该方法已应用于RBLT[33]、NARRE[2]和A3NCF[5]，并取得了良好的性能。)

• It is worth mentioning that we also tried to add a fully-connected neural layer after the fusion step. However, it leads to inferior performance duo to the overfitting problem. (值得一提的是，我们还尝试在融合步骤后添加一个完全连接的神经层。然而，由于过度装配的问题，它会导致较差的性能。)

3.4 Learning

3.4.1 Prediction.

• After completing the model training process, the recommendation task is reduced to a ranking problem among all the items in the dataset based on estimated score $\bar{R}ij$. Our prediction part is built on Matrix Factorization (MF), which is state-of-the-art for rating prediction as well as modeling implicit feedback [18]: (在完成模型训练过程后，推荐任务被简化为基于估计得分$\bar{R}ij$的数据集中所有项目的排序问题。我们的预测部分基于矩阵分解 (MF)，这是评级预测和隐式反馈建模的最新技术[18]：)
  
• The items (unclicked/ uncomsumed) are ranked in descending order of $\bar{R}ij$ to provide the Top-K item recommendation list. (项目（未点击/未消费）按$\bar{R}ij$降序排列，以提供Top-K项目推荐列表。)

3.4.2 Optimization.

• (1) Our objective is to study implicit feed-back which is more pervasive in practice and can be collected automatically (e.g. clicks, comsumes).

• To this end, we opt BPR pair-wise learning objective [26], a commonly used objective function in many previous studies [3, 34, 39–41]. (为此，我们选择BPR成对学习目标[26]，这是许多以前研究[3,34,39–41]中常用的目标函数。)

• For each positive user-item pair $<u_i,v_j>$, we randomly sample a negative item from the unobserved items of the user, which is denoted as $v_k$. (对于每个积极的用户项对$<u_i,v_j>$, 我们从用户未观察到的项目中随机抽取一个负项目，表示为$v_k$)

• The pairwise ranking loss is as follows:
  

  • where $\sigma (x)=1/(1+exp(-x))$is the logistic sigmoid function,
  • $D$ denotes the set of pairwise training instances. ($D$表示成对训练实例集。)
  • and ${\lambda }_{\Theta }$ controls the strength of regularization, which is a $L_2$ norm to prevent overfitting (和 ${\lambda }_{\Theta }$ 控制正则化的强度，这是一个$L_2$防止过拟合的项)

• (2) To optimize the objective function, we adopt mini-batch Adagrad [7] as the optimizer. Its main advantage is that the learning rate can be self-adapted during the training phase, which eases the pain of choosing a proper learning rate and leads to faster convergence than the vanilla SGD. (为了优化目标函数，我们采用mini-batch Adagrad[7]作为优化器。 它的主要优点是，学习率可以在训练阶段自适应，这减轻了选择适当学习速度的痛苦，并导致比普通SGD更快的收敛。)

4 EXPERIMENTS

4.1 Experimental Settings

4.1.1 Datasets.

• (1) In our experiments, we used three publicly accessible datasets to evaluate the performance of our model, which are Delicious1, Ciao2, and Epinions3. We briefly introduce the three datasets: (在我们的实验中，我们使用了三个可公开访问的数据集来评估我们模型的性能，它们分别是Delicious 1、Ciao2和Epinions3。我们简要介绍三个数据集：)

  • Delicious: This dataset contains social connections, book-marking, and tag information from a set of 2K users from Delicious Social Bookmarking System. In this paper we only use the social connections and book-marking records to train our model and baseline methods. (该数据集包含Delicious social Bookmarking System的2K用户的社交关系、图书标记和标签信息。在本文中，我们仅使用社会关系和簿记记录来训练我们的模型和基线方法。)
  • Ciao: This dataset contains users’ ratings to the items they have purchased and the social connections between users. Since we focus on the implicit feedback, we transform the detailed ratings into a value of 0 or 1 indicating whether the user has rated the item. (该数据集包含用户对他们购买的物品的评分以及用户之间的社会关系。由于我们关注的是隐性反馈，因此我们将详细评分转换为0或1的值，以指示用户是否对该项目进行了评分。)
  • Epinions: Epinions is a who-trust-whom directed online social network that provides product rating and review service. The corresponding rating is also assigned to a value of 1 (as implicit feedback) in our experiments. (Epinions是一个由世卫组织信任的在线社交网络，提供产品评级和评论服务。在我们的实验中，相应的评分也被指定为1（作为内隐反馈）。)

• (2) All the datasets were preprocessed to make sure that all items have at least five ratings. The statistical details of these datasets are presented in Table 1. (所有数据集都经过预处理，以确保所有项目至少有五个评分。这些数据集的统计细节见表1。)

4.1.2 Baselines.

To evaluate the performance of Top-K recommendation, we compare SAMN with the following methods. Note that all models are learned by optimizing the same pairwise ranking loss of BPR (cf Eqn. (10)) for a fair comparison. (为了评估Top-K推荐的性能，我们将SAMN与以下方法进行比较。请注意，所有模型都是通过优化相同的BPR成对排名损失（参见等式（10））来学习的，以便进行公平比较。)

• BPR [26]: This method optimizes MF with the BPR objective function. It is a highly competitive method for implicit feedback based recommendation. （该方法利用BPR目标函数对MF进行优化。对于基于隐式反馈的推荐，它是一种极具竞争力的方法。）
• SBPR [42]: This is a ranking model that considers social relationships in the learning process, assuming that users tend to assign higher ranks to items that their friends prefer. (这是一个考虑学习过程中社会关系的排名模型，假设用户倾向于为朋友喜欢的项目分配更高的排名。)
• SocialMF [15]: This is a classical model for social-aware recommendation. It incorporate the social influence among users into classical latent factor models, where the influence strength is simply set equally for all social connections. (这是一个典型的社会意识推荐模型。它将用户之间的社会影响纳入经典的潜在因素模型中，在该模型中，所有社会关系的影响强度都是相等的。)
• NCF [11]: This is a recently proposed state-of-the-art deep learning based framework that combines matrix factorization (MF) with a multilayer perceptron model (MLP) for item ranking. Since NCF shows superior performance over the traditional weighting methods WMF [14] and eALS [12] according to [11], we do not further compare with the performance of WMF and eALS. (这是一个最近提出的基于深度学习的最新框架，它将矩阵分解（MF）与多层感知器模型（MLP）相结合，用于项目排名。根据[11]，由于NCF显示出优于传统加权方法WMF[14]和eALS[12]的性能，我们不再与WMF和eALS的性能进行进一步比较。)
• SNCF: NCF [11] is designed suitable for recommendation with side information. To adjust NCF for modeling social relations, we plug user’s friends into the input feature vector and concatenate the feature vector with user ID embedding, dubbed this enhanced model as SNCF (NCF[11]的设计适合于附带信息的推荐。为了调整NCF来建模社会关系，我们将用户的朋友插入到输入特征向量中，并将特征向量与用户ID嵌入连接起来，称这种增强模型为SNCF)
• NFM [10]: This is a recently proposed Neural Factorization Machine. It is one of the state-of-the-art deep learning methods, which uses Bi-Interaction Layer to integrate both features and historical feedback information. In our experiments, we treat users’ social connections as features. Since the original NFM is designed for regression, we changed the optimize function to BPR (cf Eqn. (10)) to fit our task. (这是最近提出的一种神经因子分解机。这是最先进的深度学习方法之一，它使用 双交互层 来整合特征和历史反馈信息。在我们的实验中，我们将用户的社交关系视为特征。因为最初的NFM是为 回归 设计的，所以我们将优化函数改为BPR（参见等式（10）），以适应我们的任务。)

To the best of our knowledge, few has explored the neural networks for social-aware recommendation. Thus we compare with traditional social-aware methods SBPR and SocialMF, content-based neural method NFM, and the extended neural model SNCF. We leave out the comparison with ARSE [32], which is designed for temporal social-aware recommendation, because the performance difference may be caused by the temporal information. (就我们所知，很少有人探索神经网络用于社会意识推荐。因此，我们比较了传统的社会感知方法SBPR和SocialMF、基于内容的神经方法NFM和扩展的神经模型SNCF。我们省略了与ARSE[32]的比较，后者是为时态社会感知推荐而设计的，因为性能差异可能是由时态信息引起的。)

4.1.3 Evaluation Metrics.

• (1) To evaluate the performance of all algorithms, we calculate Recall@K and NDCG@K [21, 39]. (为了评估所有算法的性能，我们计算Recall@K和NDCG@K [21, 39])
  • Intuitively, the NDCG@K metric is a position-aware ranking metric. (直觉上NDCG@Kmetric是一种位置感知排名指标。)
  • while Recall@K metric considers whether the ground truth is ranked among the top K items. (虽然Recall@K度量标准考虑ground truth是否在前K项中排名。)
  • When K is fixed, the Precision is only determined by true positives while Recall is determined by both true positives and positive samples [39]. (当K固定时，准确度仅由真阳性决定，而召回率由真阳性和阳性样本共同决定[39]。)
• (2) To give a more comprehensive evaluation, we exhibit Recall rather than Precision and F1-score (F1-score is almost determined by Precision since Precision is much smaller than Recall in our experiments). For each user, these metrics are computed as follows: (为了给出更全面的评估，我们展示了召回率而非精确率和F1分数（F1分数几乎由精确性决定，因为在我们的实验中，精确性比召回率小得多）。对于每个用户，这些指标的计算如下：)
  
  • Where $re{l}_j=1/0$ indicates whether the item at rank $j$ in the Top-K recommendation list is in the test set, (其中，$re{l}_j=1/0$表示Top-K推荐列表中排名$j$的项目是否在测试集中，)
  • $|y_u^{test}|$ denotes the number of items rated by user $u$ in the test set. ( $|y_u^{test}|$表示用户$u$在测试集中对项目进行评分的数量。)
  • The notion IDCG means the maximum possible DCG through ideal ranking. (IDCG的概念意味着通过理想排名获得最大可能的DCG。)
  • Each metric is the average for all users. (每个指标都是所有用户的平均值。)

4.1.4 Experiments Details.

• (1) We randomly split the dataset into training (70%), validation (20%), and test (10%) sets.
• The validation set was used for tuning hyper-parameters and the final performance comparison was conducted on the test set. (验证集用于调整超参数，并在测试集上进行最终性能比较。)
• The parameters for baseline methods were initialized as in the corresponding papers, and were then carefully tuned to achieve optimal performances. (基线方法的参数在相应的论文中被初始化，然后被仔细调整以达到最佳性能。)
• The learning rate for all models are tuned amongst [0.005, 0.01, 0.02, 0.05].
• The batch size was tested in [64, 128, 256]
• and the latent factor number was tested in [32, 64, 128].
• To prevent overfitting, we turned the margin ${\lambda }_{\Theta }$ in [0.001, 0.005, 0.01, 0.02].
• For SAMN, the number of memory slices in $M$ is tuned amongst [8, 16, 32, 64].
• After the turning process, we set the latent factor numberd=128, learning rate lr=0.05, ${\lambda }_{\Theta }$ = 0.01. The memory size $N$ is set 8 for Delicious and 16 for Ciao and Epinions.
• To evaluate on different recommendation lengths, we set K = 10, 20 and 50 in our experiments.
  

4.2 Comparative Analysis on Overall Performances

• (1) The empirical results of our proposed SAMN and baselines on three datasets are given in Table 2. From the results, we can make the following observations: (表2给出了我们提出的SAMN和基线在三个数据集上的实证结果。从结果中，我们可以得出以下观察结果：)

  • Firstly, methods utilizing social information generally perform better than those without social information. For example, in Table 2, the performance of SBPR is better than BPR, SNCF performs better than NCF, and SAMN performs better than BPR and NCF. This is not surprising, since the social information is complementary to users’ historical records, it can help to increase the learning accuracy of user preference. (首先，利用社会信息的方法通常比没有社会信息的方法表现更好。例如，在表2中，SBPR的性能优于BPR，SNCF的性能优于NCF，SAMN的性能优于BPR和NCF。这并不奇怪，因为社交信息与用户的历史记录是互补的，它可以帮助提高用户偏好的学习准确性。)

  • Secondly, our method SAMN consistently and significantly out-performs all the baselines including neural methods NCF, SNCF and NFM. Specifically, SAMN outperforms the best baseline by performance gains about 6.49% on Delicious, 4.31% on Ciao and 2.43% on Epinions on the NDCG@10 metric. The performance gains on other metrics are also similarly high. (其次，我们的方法SAMN一致且显著优于所有基线，包括神经方法NCF、SNCF和NFM。具体而言，SAMN在Delicious、Ciao和Epinions上的表现分别比最佳基线提高了约6.49%、4.31%和2.43%NDCG@10米制的其他指标的性能增益也同样高。)

  • Thirdly, since SAMN share the same loss function with other baseline methods, we can attribute the performance increase to the proposed attention-based memory module and friend-level attention. In our model, the aspect-level differences and friend- level differences are considered, which allow the social information to be modeled with a finer granularity and thus lead to a better performance. (第三，由于SAMN与其他基线方法具有相同的损失函数，我们可以将性能的提高归因于所提出的基于注意的记忆模块和朋友级注意。在我们的模型中，考虑了方面级别的差异和朋友级别的差异，这使得社会信息能够以更细的粒度进行建模，从而获得更好的性能。)

• (2) We also conduct experiments to test the influence of latent factor size $d$ on validation sets. The results are shown in Figure 4. Duo to the space limitation, we only show the results of Delicious and Ciao datasets on Recall@10 and NDCG@10 metrics. (我们还进行了实验，以测试潜在因子大小dd对验证集的影响。结果如图4所示。由于空间的限制，我们只在屏幕上显示Delicious和Ciao数据集的结果Recall@10和NDCG@10韵律学。)

  • The result of Epinions is similar to that of Ciao. As can be seen from this figure, our model outperforms all the other models with different values of d for the two ranking metrics on two datasets. (Epinions的结果与Ciao的结果相似。从这个图中可以看出，对于两个数据集上的两个排名指标，我们的模型优于所有其他具有不同d值的模型。)
  • Moreover, as the latent dimension size increases, the performance of all models increase. This indicates that a larger dimension could capture more hidden factors of users and items, which is beneficial to Top-K recommendation due to the increased modeling capability. (此外，随着潜在维数的增加，所有模型的性能都会提高。这表明，更大的维度可以捕获更多用户和项目的隐藏因素，这有利于Top-K推荐，因为建模能力增强了。)

4.3 Effect of Attention-based Memory Module and Friend-level Attention

• (1) The key characteristics in our proposed model SAMN are the two newly designed components for social information modeling: (我们提出的模型SAMN的关键特征是两个新设计的用于社会信息建模的组件)

  • the attention-based memory module that captures the user-friend specific relationship, (基于注意的记忆模块，用于捕获用户与朋友之间的特定关系)
  • and the friend-level attention that models social influence strength by adaptively learning the weight of each friend. (以及通过自适应学习每个朋友的权重来建模社会影响强度的朋友级注意。)

• (2) In this subsection, we discuss the effect of attention-based memory module and friend-level attention. Specifically, we compare the effect of each component by constructing the following variants of SAMN: (在本小节中，我们将讨论基于注意的记忆模块和朋友级注意的效果。具体而言，我们通过构建以下SAMN变体来比较每种成分的效果：)

  • BPR: This is our basic collaborative filtering model without any social information. It is added as a baseline of other variants. (这是我们基本的协作过滤模型，没有任何社会信息。它是作为其他变体的基线添加的。)
  • SE: This is a variant model utilizing social information only by social embedding. (这是一种仅通过社会嵌入利用社会信息的变体模型。)
  • SAM: This is a variant model utilizing social information with the attention-based memory module only. (这是一种仅使用基于注意力的记忆模块，利用社会信息的变体模型。)
  • SFA: This is a variant model utilizing social information with the friend-level attention only. (这是一种仅利用朋友级注意力的社交信息的变体模型。)
  • SAMN: This is our proposed model utilizing social information with both the attention-based memory module and the friend-level attention. (这是我们提出的模型，利用基于注意的记忆模块和朋友级注意的社会信息。)

• The characteristics of the variant models are listed in Table 3.

• 
  

• (3) Figure 5 shows the performance of different variants. Due to the space limitation, we also only show the results of Delicious and Ciao datasets on Recall@10 and NDCG@10 metrics. (图5显示了不同变体的性能。由于篇幅限制，我们也只在屏幕上显示Delicious和Ciao数据集的结果Recall@10和NDCG@10韵律学。)

• As shown in the figure, BPR performs the worst since no social relationships are utilized to provide the extra information. (如图所示，BPR表现最差，因为没有利用社会关系来提供额外的信息。)

• With the social information of friend embedding, SE performs better, but still worse than SAM and SFA. Because just friend embedding is too crude and can provide very limited social information to help infer users’ preferences. (在嵌入好友的社交信息方面，SE的表现更好，但仍比SAM和SFA差。因为仅仅朋友嵌入太粗糙了，可以提供非常有限的社会信息来帮助推断用户的偏好。)

• The performances of SAM and SFA are significantly better than SE (p<0.05), which shows that both the attention-based memory module and the friend-level attention can help to better utilize social information by considering aspect-level differences and friend-level differences respectively. (SAM和SFA的表现明显优于SE（p<0.05），这表明基于注意的记忆模块和朋友级注意都可以通过分别考虑方面和朋友级差异来帮助更好地利用社会信息。)

• Moreover, generally SAM performs better than SFA, this may because that SAM can capture user-friend specific relation vectors through the advantage of memory network, which is more flexible than SFA that only uses the friend-level attention. (此外，SAM通常比SFA表现更好，这可能是因为SAM可以通过记忆网络的优势捕获用户朋友特定的关系向量，这比仅使用朋友级注意的SFA更灵活。)

• Lastly, our proposed SAMN, which utilizes both the attention-based memory module and the friend-level attention, performs best (significantly better than SAM and SFA for p<0.05). The reason is that these two components are not conflict with each other and can be used to model users’ social influence collaboratively. (最后，我们提出的SAMN同时利用了基于注意的记忆模块和朋友级注意，表现最好（显著优于SAM和SFA，p<0.05）。原因在于，这两个组成部分并不相互冲突，可以用来协同建模用户的社会影响力。)

4.4 Attention Analysis

• (1) The attention weights reflect how the model learned and recommend. In this subsection we conducted experiments to show how the attention-based memory module and friend-level attention component work. (注意力权重反映了模型学习和推荐的方式。在本小节中，我们进行了实验，以展示基于注意的内存模块和朋友级注意组件是如何工作的。)

• (2) We first focus on the attention-based memory module, which is designed to model the aspect-level differences between user-friend relationships. (我们首先关注基于注意力的记忆模块，该模块旨在模拟用户朋友关系之间的方面差异。)

  • Ideally, if a user-friend pair has the same interests in certain aspects, it can be reflected through the attention distributions of memory slices. (理想情况下，如果用户-朋友对在某些方面有相同的兴趣，则可以通过记忆片的注意力分布反映出来。)
  • To better understand the attention results, we make use of users’ tag information contained in Delicious dataset, which can be taken as users’ explicit preferences. (为了更好地理解注意结果，我们利用了Delicious数据集中包含的用户标签信息，这些信息可以作为用户的显式偏好。)
  • We take some examples to show how the attention weights identify different co-preferences between users and their friends in Figure 6. From the figure, we can see that for user-friend pair ($u_x$, $u_a$) and ($u_x$, $u_b$), the attention distributions of memory slices are similar, which means that $u_a$ and $u_b$ share similar interest aspects with $u_x$. Since they all have the tag “sport”, we can imagine that the shared interest aspects may be sports-related matters. (我们举一些例子来说明注意权重如何识别用户和他们的朋友在图6中的不同共同偏好。从图中，我们可以看到，对于用户朋友对（$u_x$，$u_a$）和（$u_x$，$u_b$），内存片的注意力分布是相似的，这意味着$u_a$和$u_b$与$u_x$有相似的兴趣方面。由于它们都有“运动”的标签，我们可以想象，共同的利益方面可能是与运动有关的事情。)
    • For pair ($u_x$, $u_c$), the distribution is different from the above two pairs. The reason may because that the shared interest aspects of $u_x$ and $u_c$ are games, not sports. With the help of attention mechanisms, our model can capture users’ attention weights on different aspects of each friend, and thus could achieve more accurate predictions (cf. Section. 4.3). (对于pair（$u_x$，$u_c$），分布与上述两对不同。原因可能是因为$u_x$和$u_c$的共同利益是游戏，而不是体育。在注意机制的帮助下，我们的模型可以捕捉用户对每个朋友不同方面的注意权重，从而实现更准确的预测（参见第4.3节）。)

• (3) Apart from the attention-based memory module, another key advantage of SAMN is its ability in adaptively measuring the influence strength of users’ friends.

  • To show this, we randomly selected a user who has four friends (#782, #1391, #1446, and #1505) in social network form Ciao dataset. We then randomly picked three items (#130, #212, and #1258) which have been purchased by the user, and three negative items (#29, #1105, and #3367) which have not been purchased.
  • Table 4 shows the attention weights of the user’s friends for the randomly selected items.
  • 
  • We have the following observations:
    • (1) For different item, the attention weights of the user’s friends vary significantly. For example, when predicting the user’s preference on item #1258, the attention weight of friend #782 are relatively high. The reason may because that the friend #782 has purchased item #1258 (according to the dataset), and thus he has a bigger influence strength for the user’s purchasing behavior on this item. (对于不同的项目，用户朋友的注意力权重有显著差异。例如，当预测用户对项目#1258的偏好时，朋友#782的注意力权重相对较高。原因可能是因为朋友#782购买了#1258物品（根据数据集），因此他对用户购买该物品的行为有更大的影响力。)
    • (2) The attention weights can also somehow reflect the richness of friend users’ feedback information. For example, the friend #1391 has only purchased 2 items according to the dataset. For both positive items and negative items, his attention weights are relatively low, which means that his information is not rich enough to provide significant influence for the user’s purchasing behaviors. (注意权重也可以在某种程度上反映朋友用户反馈信息的丰富程度。例如，根据数据集，朋友#1391只购买了2件物品。对于积极项目和消极项目，他的注意力权重都相对较低，这意味着他的信息不够丰富，无法对用户的购买行为产生显著影响。)

5 CONCLUSION

• (1) Social information plays a very important role for improving the performances of recommender systems. (社会信息对于提高推荐系统的性能起着非常重要的作用。)
  • However, there are two differences: aspect-level differences among user-friend co-preferences and friend-level differences on social influence strength, have not been well-studied in existing methods. (然而，存在两个差异：用户朋友共同偏好之间的方面水平差异和社会影响力强度上的朋友水平差异，在现有方法中尚未得到很好的研究。)
• (2) In this paper, we proposed a new model, which unifies the strengths of memory networks and ==attention mechanisms ==to address the problems in social-aware recommendations. (在本文中，我们提出了一个新的模型，该模型结合了记忆网络和注意力机制的优点，以解决社会意识推荐中的问题。)
  • To the best of our knowledge, we are the first to employ the attention-based memory module to construct user-friend specific relation vector. (据我们所知，我们是第一个使用基于注意的记忆模块来构建用户朋友特定关系向量的人。)
  • We also design the friend-level attention to adaptively measure the social influence strength among users’ friends. (我们还设计了朋友级注意，以自适应地测量用户朋友之间的社会影响力强度。)
• (3) Extensive experiments have been made on three real-life datasets. The proposed SAMN (Social Attentional Memory Network) consistently and significantly outperforms the state-of-the-art recommendation models on different evaluation metrics.
  • Moreover, we performed qualitative analyses of the attentional memory module and the friend-level attention, which helps understand what the model has learnt and prove the rationality of our model. (在三个真实数据集上进行了广泛的实验。所提出的SAMN（社会注意记忆网络）在不同的评估指标上持续且显著优于最先进的推荐模型。)
• (4) In the future, we are interested in exploring the differences between the two types of social connections: friends in undirected social networks and followers in directed social networks. Moreover, since the social information can be used to explain the recommendation results, we also like to investigate and improve the explainability of our model. (在未来，我们有兴趣探索两种类型的社会关系之间的差异：无定向社交网络中的朋友和定向社交网络中的追随者。此外，由于社会信息可以用来解释推荐结果，我们也希望调查并提高我们模型的可解释性。)

ACKNOWLEDGMENTS

REFERENCES