[论文阅读笔记]2020_CIKM_DREAM_ A Dynamic Relation-Aware Model for social recommendation

论文下载地址： https://doi.org/10.1145/3340531.3412115
发表期刊：CIKM
Publish time: 2020
作者及单位:

Liqiang Song, Ye Bi, Mengqiu Yao, Zhenyu Wu, Jianming Wang, Jing Xiao
magicyebi@163.com,{songliqiang537,yaomengqiu621,wuzhenyu447,wangjianming888,xiaojing661}@pingan.com.cn
Ping An Technology Shenzhen Co., Ltd

数据集： 正文中的介绍

Epinions http://www.epinions.com （作者没给）
Douban http://movie.douban.com (作者自己爬的)
代码：
（作者没给）

其他：

其他人写的文章

简要概括创新点：考虑了temporal，TIE(涉及到session) + GAT + GloVe-based method。感觉把NLP里的模型，往SoRec套。

We propose a novel RS approach, which aims to model users’ dynamic interests and dynamic influences from their friends. (我们提出了一种新的RS方法，旨在模拟用户的动态兴趣和朋友的动态影响。)

The model encodes the outputs form historical sessions by recursively combining the features encoded by relational-GAT modules and that from last TIE module. (该模型通过递归组合关系GAT模块和最后一个TIE模块编码的特征，对历史会话的输出进行编码。)

We design a GloVe-based method to increase the number of friends, and use relational-GAT to aggregate user representations from both completed social network in each session. (我们设计了一种基于GloVe的方法来增加朋友的数量，并使用关系GAT来聚合每个会话中来自两个完整社交网络的用户表示。)

细节

Epinions. users’ behaviors are segmented into month-long sessions. (用户的行为被分割为期一个月的（以月为单位的）sessions。)

DouBan. segment users’ behaviors into week-long sessions for their high activeness. (将用户的行为划分为为期一周的(以周为单位的)会议，以提高他们的活跃度)
We convert users’ explicit ratings on items into 1 as the implicit feedback for all the datasets. (我们将用户对项目的显式评分转换为1，作为所有数据集的隐式反馈。)

We use Adam [5] as the optimizing method for all models that relied on the gradient descent based methods

In order to make the models converge faster, we apply the batch normalization. (为了使模型更快地收敛，我们采用了批量归一化。)

ABSTRACT

(1) Social connections play a vital role in improving the performance of recommendation systems (RS). However, incorporating social information into RS is challenging. (社会关系在提高推荐系统（RS）的性能方面起着至关重要的作用。然而，将社会信息整合到RS中是一个挑战。)
- Most existing models usually consider social influences in a given session, ignoring that both users’ preferences and their friends’ influences are evolving. (大多数现有的模型通常考虑在给定的会话中的社会影响，忽略了用户的偏好和他们的朋友的影响正在演变。)
- Moreover, in real world, social relations are sparse. Modeling dynamic influence and alleviating data sparsity is of great importance. (此外，在现实世界中，社会关系很少。建模动态影响和缓解数据稀疏性非常重要。)
(2) In this paper, we propose a unified framework named Dynamic RElation-Aware Model (DREAM) for social recommendation, which tries to model both users’ dynamic interests and their friends’ temporal influences. (在本文中，我们提出了一个统一的社会推荐框架，称为动态关系感知模型（DREAM），它试图同时模拟用户的动态兴趣和他们朋友的时间影响。)
- Specifically, we design temporal information encoding modules, because of which user representations are updated in each session. The updated user representations are transferred to relational-GAT modules, subsequently influence the operations on social networks. (具体来说，我们设计了时态信息编码模块，因为每个会话中都会更新用户表示。更新后的用户表示被转移到关系GAT模块，从而影响社交网络上的操作。)
- In each session, to solve social relation sparsity, we utilize glove-based method to complete social network with virtual friends. (在每个会话中，为了解决社交关系稀疏问题，我们使用基于手套的方法来完成与虚拟朋友的社交网络。)
- Then we employ relational-GAT module over completed social networks to update users’ representations. (然后，我们在完整的社交网络上使用关系GAT模块来更新用户的表示。)
(2) In the extensive experiments on the public datasets, DREAM significantly outperforms the state-of-the-art solutions. (在对公共数据集进行的大量实验中，DREAM的表现明显优于最先进的解决方案。)

CCS CONCEPTS

• Applied computing → Online shopping; • Networks → Online Social Networks.

KEYWORDS

Session-based Social Recommendation, Virtual Friends, Temporal Information Encoding

1 INTRODUCTION

(1) Recommendation systems (RS), attempting to help users overcome information overload, have become popular. RS make use of users’ historical activities and/or social relationships to generate features of items and identify users’ latent preferences. Among RS strategies, collaborate filtering (CF)-based methods have received significant success [7]. However, CF-based methods usually suffer from sparsity of user-item interactions and cold start problem. (试图帮助用户克服信息过载的推荐系统（RS）已经很流行。RS利用用户的历史活动和/或社会关系来生成项目的特征，并识别用户的潜在偏好。在RS策略中，基于 协同过滤（CF）的方法 取得了显著的成功[7]。然而，基于CF的方法通常存在用户项交互稀疏和冷启动问题。)
(2) To address these limitations, modern RS also generate other useful data. By combining rating matrices with additional data, better recommendations are respected. Nowadays, the increasing popularity of social media greatly enriches people’s social activities, which harnesses social relations to boost the performance of RS [1]. Based on the intuition that people in same social group are likely to have similar preferences, and that users will gather information from their social friends, lots of work has incorporated social information into RS, e.g. SBPR [11], GraphRec [1], etc…However, incorporating social information into RS is challenging. (为了解决这些限制，现代遥感器还生成了其他有用的数据。通过将评级矩阵与其他数据相结合，更好的建议得到了尊重。如今，社交媒体的日益普及极大地丰富了人们的社交活动，利用社交关系提升了RS的表现[1]。基于这样一种直觉，即同一社会群体中的人可能有相似的偏好，并且用户会从他们的社交朋友那里收集信息，许多工作已经将社交信息纳入了RS，例如 SBPR [11]，GraphRec [1]等。然而，将社交信息纳入RS是一个挑战。)
(3) In general, methods utilizing heterogeneous data tend to perform better than those using a single data source [3]. (一般来说，使用异构数据的方法往往比使用单一数据源的方法性能更好[3]。)
- DGRec [8] models dynamic user behaviors with RNN, and social influence with GAT, and is proven to outperform the state-of-the-art solutions. (DGRec[8]使用RNN模拟动态用户行为，使用GAT模拟社会影响，并被证明优于最先进的解决方案。)
- However, the work only considers social influences in a given session, ignoring the effects of temporal dependency among different sessions. (然而，这项工作只考虑了给定会话中的社会影响，忽略了不同会话之间的时间依赖性的影响。)
- Besides, relying only on social friends is far from satisfactory, since in real world, the social relations are very sparse. (此外，仅仅依靠社交朋友是远远不能令人满意的，因为在现实世界中，社交关系非常稀少。)
(4) Based on the observations, we propose a novel framework called A Dynamic RElational-Aware Model (DREAM), which tries to model both users’ dynamic individual interests and their friends’ temporal influences. (基于这些观察，我们提出了一个称为动态关系感知模型（DREAM）的新框架，该框架试图对用户的动态个人兴趣和他们朋友的时间影响进行建模。)
The model contains two main modules, (该模型包含两个主要模块)
- temporal information encoding (TIE) modules encode the outputs form historical sessions by recursively combining the representation of all previous outputs with the current output. (时间信息编码（TIE）模块通过递归地将所有先前输出的表示与当前输出相结合，对来自历史会话的输出进行编码。)
- In each session, we first complete social network, and then employ relational-aware graph attention (relational-GAT) network modules to integrate influences from users’ real and virtual friends. ( 在每个会话中，我们首先完成社交网络，然后使用关系感知图注意（relational GAT）网络模块来整合用户 真实和虚拟朋友的影响。)
To be specific, we first design a method to increase the number of friends, which refer to the neighbors whose behaviors are similar to the users, and we call this kind of neighbors as virtual friends. (具体来说，我们首先设计了一种增加朋友数量的方法，即行为与用户相似的邻居，我们称这种邻居为虚拟朋友。)
- Within each session, we then utilize relational-GAT [10] to capture the influences of friends. 在每个环节中，我们利用关系GAT[10]来捕捉朋友的影响。
- As we know, user preference for products drifts over time.
- To model users’ dynamic preferences and the dynamic influences from their friends, we design the temporal information encoding modules. (为了模拟用户的动态偏好和朋友的动态影响，我们设计了时态信息编码模块。)
- In each TIE module, we combine the relational-GAT encoded features as well as the features from last TIE module.(在每个环节中，我们利用关系GAT[10]来捕捉朋友的影响。正如我们所知，用户对产品的偏好会随着时间的推移而变化。为了模拟用户的动态偏好和朋友的动态影响，我们设计了时态信息编码模块。在每个TIE模块中，我们结合了关系GAT编码的特征以及来自上一个TIE模块的特征)
(5) In summary, our contributions in this paper are as follows: (总之，我们在本文中的贡献如下：)
- We propose a novel RS approach, which aims to model users’ dynamic interests and dynamic influences from their friends. (我们提出了一种新的RS方法，旨在模拟用户的动态兴趣和朋友的动态影响。)
- The model encodes the outputs form historical sessions by recursively combining the features encoded by relational-GAT modules and that from last TIE module. (该模型通过递归组合关系GAT模块和最后一个TIE模块编码的特征，对历史会话的输出进行编码。)
- We design a GloVe-based method to increase the number of friends, and use relational-GAT to aggregate user representations from both completed social network in each session. (我们设计了一种基于GloVe的方法来增加朋友的数量，并使用关系GAT来聚合每个会话中来自两个完整社交网络的用户表示。)
- We conduct experiments on real-world recommendation scenarios, and the results prove the efficacy of DREAM over several state-of-the-art baselines. (我们在现实世界的推荐场景上进行了实验，结果证明了DREAM在几个最先进的基线上的有效性。)

2 PROBLEM FORMULATION

Let $U={u1,u2,...un}\mathcal{U} = \{u_1, u_2, . . .u_n\}$ and $I={i1,i2,...im}\mathcal{I} = \{i_1, i_2, . . .i_m\}$ denote the sets of users and items, (表示用户和项目的集合，)
where $n$ and $m$ are the number of users and items, respectively. (其中， $n$ 和 $m$ 分别是用户和项目的数量。)
The user-item interaction matrix $\in R^{n×m}$ is defined as: $y_{u,i} = 1$ , if there is an interaction between $u$ and $i$ , and 0 otherwise.
Each $u$ is associated with a set of sessions, $ITu={S1u,S2u,...,STu}\mathcal{I}^u_T= \{S^u_1,S^u_2, . . . ,S^u_T\}$ , (每个 $u$ 都与一组会话相关联)
where $StuS^u_t$ is the $t$ -th session of user $u$ , (是用户 $u$ 的第 $t$ 次会话，)
each session $StuS^u_t$ consists of a user behaviors sequence ${it,1u,it,2u,...,it,Nu,tu}\{i^u_{t,1}, i^u_{t,2}, . . . , i^u_{t,N_{u,t}}\}$ , (每个会话 $StuS^u_t$ 由一个用户行为序列组成)
where $it,pui^u_{t,p}$ is the $p$ -th item interacted by user $u$ in $t$ -th session. (是用户 $u$ 在第 $t$ 次会话中交互的第 $p$ 次项。)
Social network can be described by $SR∈Rn×nS^R \in R^{n×n}$ , where $sp,qR=1s^R_{p,q} = 1$ if there is a relation between $u_p$ and $u_q$ , and 0 otherwise.
We call embedding vectors $u$ and $i$ latent feature vectors. (我们称嵌入向量为 $u$ 和 $i$ 潜在特征向量。)
Given $Y$ and $S^R$ , we aim to predict whether $u$ has potential interests in target item $v$ .

3 DREAM FRAMEWORK

(1) The framework is illustrated in Figure 1.
(2) The model consists of
- social network completion, (社交网络补全)
- relational-aware graph attention network (relational-GAT) modules, (关系感知图形注意网络（relational GAT）模块)
- temporal information encoding (TIE) modules (时间信息编码（TIE）模块)
- and recommendation. (正式建议)
(3) Given a user $u$ , we first select her $T$ historical sessions. (给定一个用户 $u$ ，我们首先选择她的 $T$ 历史会话。)
- In $t$ -th session, we first complete social network, and get $GtC\mathcal{G}^C_t$ . The inputs of the model are graph sequence ${G1C,G2C,...,GTC}\{\mathcal{G}^C_1, \mathcal{G}^C_2, . . . , \mathcal{G}^C_T\}$ . (在 $t$ -th会话中，我们首先完成(补全)社交网络，并获得 $GtC\mathcal{G}^C_t$ 。模型的输入是图形序列 ${G1C,G2C,...,GTC}\{\mathcal{G}^C_1, \mathcal{G}^C_2, . . . , \mathcal{G}^C_T\}$ 。)
(4) In each session, relational is employed to integrate influences from $GtC\mathcal{G}^C_t$ . (在每一个会话中，关系被用来整合来自 $GtC\mathcal{G}^C_t$ 的影响。)
(5) User representations are updated in TIE module and transferred to next relational module. (用户表示在TIE模块中更新，并传输到下一个关系模块。)

3.1 Social Network Completion

3.1.1 Virtual Friends Definition and Selection.

(1) We define virtual friend as users having similar consumption habits, and the connection is stronger if they are more similar. (我们将虚拟朋友定义为具有相似消费习惯的用户，如果他们更相似，联系就会更强。)
(2) Under the definition, we design a GloVe-based method. (根据定义，我们设计了一种基于GloVe的方法)
- We first utilize GloVe mechanism [6] to learn user representations, (我们首先利用GloVe机制[6]来学习用户表示，)
- and then calculated the similarity among all users, (然后计算所有用户之间的相似度，)
- finally we choose top-k users whose similarity is higher. (最后选择相似度较高的top-k用户。)
- The input is user-user co-occurrence counts matrix $X$ , (输入为用户共现计数矩阵 $X$ )
  - whose entries $X_{p,q}$ denote the number of times user $p$ and user $q$ consume the same items. (条目 $X_{p，q}$ 表示用户 $p$ 和用户 $q$ 消费相同项目的次数。)
  - And outputs are user embeddings $g_u$ . (输出是用户嵌入的 $g_u$ .)
(3) We define the connection strength between user $u^{p}$ and $u^{q}$ as: (我们将用户 $u^{p}$ 和 $u^{q}$ 之间的连接强度定义为：)
- and $u_p$ ’s virtual friends set is defined as $N^V(u_p)$ . ( $u_p$ 的虚拟朋友集定义为 $N^V(u_p)$ 。)

3.1.2 Node Representation.

(1) Friends’ influences always lag, since friends may consumed products first and then influence the user. So in $t$ -th relational-GAT module, friends node representations are calculated in $(t - 1)$ -th session. (朋友的影响力总是滞后的，因为朋友可能会先消费产品，然后影响用户。因此，在第 $t$ -th关系GAT模块中，朋友节点表示在第 $(t - 1)$ -th会话中计算。)
(2) Here, we use users’ short-term interests, which are gotten by employing GRU [2] over user $u^{j}$ ’s interaction sequence in $(t - 1)$ -th session $St−1j={it−1,1j,it−1,2j,...,it−1,Nk,tj}S^j_{t−1}= \{ i^j_{t−1,1}, i^j_{t−1,2}, . . . , i^j_{t−1,N_{k,t}} \}$ $S^{t - 1 j} = {i^{t - 1, 1 j}, i^{t - 1, 2 j}, . . ., i^{t - 1, N^{k, t} j}}$ , (在这里，我们使用用户的短期兴趣，这是通过在用户 $u^{j}$ 的第 $(t - 1)$ 次会话的交互序列上使用GRU[2]得到的)
- i.e. $sj=GRU(St−1j)s_j = GRU(S^j_{t−1})$ .

3.2 Relational-GAT Module

(1) Note there are two different relations in our completed social network, to capture this difference, we employ relational-GAT over the completed social network in each session. (注意在我们完整的社交网络中有两种不同的关系，为了捕捉这种差异，我们在每个会话中对完整的社交网络使用关系GAT。)
(2) In $t$ -th session, for each user, we build a graph where nodes correspond to target user and her friends. (在 $t$ -th会话中，我们为每个用户构建一个图，其中节点对应于目标用户及其朋友。)
- For target user $u$ with $∣ N (u) ∣$ friends, the graph has $∣ N (u) ∣$ + 1 nodes. (对于目标用户 $u$ 和 $∣ N (u) ∣$ 朋友，该图有 $∣ N (u) ∣$ + 1个节点。)
- we rerepresent nodes as $hu(0)=u~t−1h^{(0)}_u = \tilde{u}_{t−1}$ $h^{u (0)} = u ~^{t - 1}$ and ${hj(0)=sj}\{ h^{(0)}_j = s_j \}$ ${h^{j (0)} = s^{j}}$ ,
  - where user node representation $u~t−1\tilde{u}_{t−1}$ is gotten from last TIE module (see section 3.3). (其中用户节点表示 $u~t−1\tilde{u}_{t−1}$ 来自上一个TIE模块（见第3.3节）。)
- Then we employ relational-GAT over $N (u)$ . (然后我们在 $N (u)$ 上应用关系GAT。)
Specifically, we first project friends to the same space, and then calculate the attention score: (具体来说，我们首先将朋友投射到同一空间，然后计算注意力得分)
- where $f_r(·,·)$ is the deep neutral network performing relational attention. (是一个深度神经网络，负责进行关系注意力。)
(3) Then, we aggregate information from $N (u)$ : (然后，我们从 $N (u)$ 聚合信息：)
- where $σ\sigma$ denotes the activation function. (其中 $σ\sigma$ 表示函数。)
- We denote the final representation of user as $u_t = h_u$ . (我们将用户的最终表示形式表示为 $u_t = h_u$ 。)

3.3 Temporal Information Encoding Module

(1) Temporal information is an important factor and there are traditional RS that consider temporal information [8]. (时间信息是一个重要的因素，传统的RS考虑时间信息。)
However, exploiting temporal personal preference and dynamic influences from users’ friends together is still challenging. (然而，同时利用用户的暂时个人偏好和朋友的动态影响仍然具有挑战性)
- For example, traditional RNN usually performs worse as the length of the sequences increases. (例如，随着序列长度的增加，传统RNN的性能通常会变差。)
- Besides in the real world, the influences produced in the earlier session may decay as time going by. (此外，在现实世界中，早期会议产生的影响可能会随着时间的推移而减弱。)
(2) Inspired by Gated Recurrent Unit (GRU) [2] we propose the temporal information encoding modules. (受门控循环单元（GRU）[2]的启发，我们提出了时态信息编码模块。)
As shown in Figure 1, we first select users’ historical behavior sessions, and utilize temporal information encoding modules among the outputs from each session to grasp the dynamic interests and influences. （如图1所示，我们首先选择用户的历史行为会话，并在每个会话的输出中使用时间信息编码模块来抓取动态兴趣和影响。）
The output of relational-GAT module in sessiont is denoted as $u_t$ . (sessiont中关系GAT模块的输出表示为 $u_t$ 。)
To combine relational-GAT encoded features at each time, as well as the target user’s dynamic personal interests, we design a GRU-like module, called TIE module. (为了将每次的关系GAT编码特征以及目标用户的动态个人兴趣结合起来，我们设计了一个类似GRU的模块，称为TIE模块。)
The hidden layer of TIE is a linear interpolation between the last TIE’s hidden layer $u~t−1\widetilde{u}_{t−1}$ $u^{t - 1}$ and the candidate hidden layer $ht~\widetilde{h_t}$ $h^{t}$ . The encoding procedures are defined as below: (TIE的隐藏层是上一条TIE的隐藏层 $u~t−1\widetilde{u}_{t−1}$ $u^{t - 1}$ 和候选隐藏层 $ht~\widetilde{h_t}$ $h^{t}$ 的一个线性插值。编码过程定义如下：)
- where $WqtW^t_q$ , $WetW^t_e$ , $WhtW^t_h$ , $Uht∈Rd×dU^t_h \in R^{d×d}$ , $bqt,bet,bht∈Rdb^t_q, b^t_e, b^t_h \in R^d$ .
- We let $u~0=u\widetilde{u}_0 = u$ , the initial user latent feature $u$ . (初始用户潜在功能 $u$ 。)
- So the vector $u~t\widetilde{u}_t$ can be seen as user’s long-term preference combining her evolving interests and the dynamic influences from her friends. (因此，向量 $u~t\widetilde{u}_t$ 可以被视为用户的长期偏好结合了用户不断发展的兴趣和来自朋友的动态影响。)

3.4 Prediction

(1) The final output of relational-GAT module == $u_T$ = is used as the target user representation, (关系GAT模块 $u_T$ 的最终输出用作目标用户表示)
(2) then the user representation $u^{T}$ and the target item $v$ are combined to predict the clicking probability $y^uv=σ(f(uT,v))\hat{y}_{uv} = \sigma (f (u_T,v))$ $y^^{u v} = σ (f (u^{T}, v))$ , (然后将用户表示 $u^{T}$ 和目标项 $v$ 结合起来，预测点击概率)
- where $σ (\cdot, \cdot)$ is sigmoid function,
- and $f$ is a ranking function which can be either a dot-product function or a deep neural network. ( $f$ 是一个排序函数，可以是点积函数，也可以是深度神经网络。)
- The loss function $L$ is the sigmoid cross entropy loss:

4 EXPERIMENT

To comprehensively study our proposed model DREAM, we conduct experiments on real-world datasets to answer the following questions: (为了全面研究我们提出的模型 DREAM，我们在现实世界的数据集上进行实验，以回答以下问题：)

Q1: Does DREAM outperform the state-of-the-art base-lines? (问题1：DREAM是否超越了最先进的基线？)
Q2: How is DREAM affected by each component? (问题2：DREAM是如何受到每个组成部分的影响的？)

4.1 Experimental Setup

4.1.1 Datasets.

(1) Epinions1and Douban-Movie2(Short for Movie) are utilized to evaluate the performance of DREAM. (Epinions1和豆瓣电影（电影的缩写）被用来评估梦的表现。)
(2) As for Epinions, “trust” endorsements are converted to directed edges in social network (至于Epinions，“信任”代言在社交网络中被转化为定向边缘)
- and users’ behaviors are segmented into month-long sessions. (用户的行为被分割为期一个月的sessions。)
(3) As for Douban-Movie , we crawled the interaction data using the identities of users in movie community along with associated timestamps and users’ social netowrks. (对于豆瓣电影，我们使用电影社区中用户的身份以及相关的时间戳和用户的社交网络标记来抓取交互数据。)
- We construct our datasets by using each review as an evidence that a user consumed an item (我们通过使用 每次评论 作为用户消费物品的证据来构建数据集)
- and segment users’ behaviors into week-long sessions for their high activeness. (将用户的行为划分为为期一周的会议，以提高他们的活跃度)
- We convert users’ explicit ratings on items into 1 as the implicit feedback for all the datasets. (我们将用户对项目的显式评分转换为1，作为所有数据集的隐式反馈。)
(4) We randomly split the user-item interactions of each dataset into training set (80%) to learn the parameters, validation set (10%) to tune hyper-parameters, and testing set (10%) for the final performance comparison. The statistics of two datasets are presented in Table 1. (我们将每个数据集的用户项交互随机分为训练集（80%）学习参数，验证集（10%）调整超参数，测试集（10%）进行最终性能比较。两个数据集的统计数据如表1所示。)

4.1.2 Baselines.

(1) In this subsection, we compare DREAM with four groups of recommendation baselines. (在本小节中，我们将DREAM与四组推荐基线进行比较。)
- (1) only considers user feedbacks: BPR [7]. (仅考虑用户反馈：BPR[7]。)
- (2) considers social network information: GraphRec [1] and SBPR [11]. (考虑社交网络信息：GraphRec[1]和SBPR[11]。)
- (3) considers the sequence of user actions: GRU [2] and SASRec [4]. (考虑用户操作的顺序：GRU[2]和SASRec[4]。)
- (4) combines social network and user actions’ temporal information into recommendation task: DGRec [? ]. (将社交网络和用户行为的时间信息合并到推荐任务中：DGRec[？]。)
(2) We also construct several variants of DREAM for ablation study on two datasets. There are two special changes in our DREAM. (我们还在两个数据集上构建了用于消融研究的DREAM的几种变体。我们的DREAM有两个特别的变化。)
- One is the addition of virtual friends’ information in social network, (一个是在社交网络中添加虚拟朋友的信息，)
- the other change is the utilization of more than one session information (the best parameter is 2). (另一个变化是利用了多个会话信息（最佳参数为2）。)
- As for inner-session information, we design two variants of DREAM, DREAM-R and DREAM-V to evaluate the effect of virtual friends’ information. (对于内部会话信息，我们设计了DREAM的两个变体DREAM-R和DREAM-V来评估虚拟朋友信息的效果。)
  - DREAM-R removes virtual friends’ information, (DREAM-R删除虚拟朋友的信息，)
  - while DREAM-V uses virtual friends’ information only. (DREAM-V只使用虚拟朋友的信息。)
- As for inter-session information, (至于会话间信息)
  - DREAM-GAT utilizes GAT[9] to combine two friend-level information rather than realtional-GAT. (DREAM-GAT利用GAT[9]将两个朋友级别的信息结合起来，而不是真实的GAT。)
  - And, DREAM-TGRU leverages GRU to fuse two session information to evaluate the necessity of TIE module. (DREAM-TGRU利用GRU融合两个会话信息来评估TIE模块的必要性。)
  - DREAM-1 uses only one session’s information like DGRec, DREAM-3 utilizes 3 sessions’ information. (DREAM-1只使用一个会话的信息，如DGRec，DREAM-3使用三个会话的信息。)

4.1.3 Metrics.

R@K (Short for Recall@K,K=10), NDCG and MRR are utilized to measure the performance. (R@K（缩写为Recall@K，K=10），NDCG和MRR用于测量性能。)
For all the metrics, the larger the values, the better the performance. (对于所有指标，值越大，性能越好。)
In order to reduce the computational since there are too many unrated items, we randomly sample 1000 unrated items as negative samples and combine them with positive items in the ranking process for each user. (为了减少计算量，因为有太多的未分级项目，我们随机抽取1000个未评分的项目作为负样本，并在每个用户的排名过程中将其与正项目相结合。)
We repeat this procedure 10 times and report the average ranking results. (我们重复这个过程10次，并报告平均排名结果。)

4.1.4 Parameter Setting.

(1) For all the models that are based on the latent factor models, we initialize the latent vectors with small random values. (对于所有基于潜在因子模型的模型，我们用小的随机值初始化潜在向量。)
We use Adam [5] as the optimizing method for all models that relied on the gradient descent based methods with a (我们使用Adam[5]作为所有模型的优化方法，这些模型依赖于基于梯度下降的方法，具有)
learning rate of 0.0001
and batch size of 32.
(2) In DREAM model, GloVe algorithm used to find high-quality virtual friends gets the best performance compared to rating matrix. (在DREAM模型中，用于寻找高质量虚拟朋友的GloVe算法得到了比评分矩阵更好的性能。)
The real and virtual friends sampling numbers are uniformly set to 10 after lots of experiments. (经过大量实验，真实朋友和虚拟朋友的抽样数被统一设置为10。)
(3) For relational-GAT module in DREAM, we set the corresponding parameters according to their original paper [10]. (对于DREAM中的关系GAT模块，我们根据他们的原始论文[10]设置了相应的参数。正则化参数为0.00001)
The regularization parameter as 0.00001
and ReLU is utilized to implement the non-linear transformation function. (利用ReLU实现非线性变换函数。)
In order to make the models converge faster, we apply the batch normalization. (为了使模型更快地收敛，我们采用了批量归一化。)
The best models are selected by early stopping when the validation accuracy does not increase for 5 consecutive epochs. (当验证精度在连续5个时间段内没有增加时，通过提前停止来选择最佳模型)
There are several other parameters in the baselines, we tune all these parameters to ensure the best performance of the baselines for fair comparison. (基线中还有其他几个参数，我们调整所有这些参数以确保基线的最佳性能，以便进行公平比较。)

4.2 Model Analysis

4.2.1 Overall Performance: Q1.

(1) As shown in Table 2, the following observations can be made: (如表2所示，可以进行以下观察)
- SBPR and GraphRec utilize both user-item interactions and social relations: (SBPR和GraphRec利用了用户项交互和社会关系：)
- while BPR only uses user-item interactions. (而BPR只使用用户项交互。)
- SBPR and GraphRec outperform BPR, which is consistent with previous work. (SBPR和GraphRec的表现优于BPR，这与之前的工作一致。)
- This indicates that social information reflects users’ interests effectively. (这表明社交信息有效地反映了用户的兴趣)
(2) In most cases, GRU and SASRec obtain better performance than BPR, which are modeled with temporal information. These improvements reflect the power of temporal information on RS. (在大多数情况下，GRU和SASRec比BPR获得更好的性能，BPR是用时态信息建模的。这些改进反映了时间信息对遥感的影响。)
- However, the performance of methods of social relations sometimes exceed temporal information based methods, which suggests it is hard to determine which is better. (然而，社会关系方法的性能有时超过基于时间信息的方法，这表明很难确定哪种方法更好。)
- DGRec combining social relations and temporal information achieves much better performance than classic, social and temporal baselines. (DGRec结合了社会关系和时态信息，比经典的、社会的和时态的基线实现了更好的性能。)
- DGRec utilizes dynamic interests and static interests social information to represent target user’s interests, meanwhile, it employs graph-based algorithm with attention mechanism to deal with the influence of social friends, which shows the effectiveness of these two changes. (DGRec利用动态兴趣和静态兴趣的社会信息来表示目标用户的兴趣，同时采用基于图的算法和注意机制来处理社交朋友的影响，这两种变化的有效性得到了体现。)
(3) Our model DREAM consistently outperforms all the baselines. The substantial improvement of DREAM over the baselines could be attributed to three reasons: (我们的DREAM模型始终胜过所有的基线。与基线相比，DREAM的显著改善可归因于三个原因：)
- (1) We expand target user’s social network using virtual friends, which comprehensively expresses target user’s dynamic and static interests. (我们利用虚拟朋友扩展目标用户的社交网络，全面表达目标用户的动态和静态兴趣。)
- (2) We combine user’s historical representation and current representation as the input of TIE. Utilizing updated user representation considers is the important step to learn the evolution of target user’s interests. (我们结合用户的历史表示和当前表示作为TIE的输入。利用更新的用户表示是了解目标用户兴趣演变的重要步骤。)
- (3) Different from DGRec only using one session information, we design TIE module to employ multiple temporal sessions’ information which reflects the evolution of target user’s dynamic interests over time. (与只使用一个会话信息的DGRec不同，我们设计了TIE模块来使用多个时态会话的信息，这些信息反映了目标用户的动态兴趣随时间的变化。)
- This strongly indicates the advantage of using temporal information across sessions. In terms of time complexity, we find that DREAM using 2-session information (25 seconds) increases 1.9 seconds than DGREC (23.1 seconds) by calculating running time spent on single GPU per epoch. (这强烈表明了跨会话使用时间信息的优势。在时间复杂度方面，我们发现，通过计算每个epoch在单个GPU上花费的运行时间，使用2个会话信息（25秒）的DREAM比DGREC（23.1秒）增加1.9秒。)

4.2.2 Ablation Study: RQ2.

(1) Table 3 are the ablation study’s results of DREAM in terms of R@10 and MRR. (表3是消融研究的DREAM境结果R@10还有MRR。)
For inner-session aspect, we compare DREAM with DREAM-R and DREAM-V. (在内部会话方面，我们将DREAM与DREAM-R和DREAM-V进行比较)
- DREAM-V captures users’ dynamic interest information hidden in graph structure, (DREAM-V捕捉隐藏在图形结构中的用户动态兴趣信息，)
- while DREAM-R makes up corresponding social information. (而DREAM-R则构成了相应的社会信息。)
- So, DREAM using both real and virtual friends’ information gets better performance than DREAM-R and DREAM-V. (因此，使用真实和虚拟朋友信息的DREAM比DREAM-R和DREAM-V性能更好。)
In inter-session aspect, we also evaluate the necessity of relational-GAT dealing with differernt kinds of social information and TIE module. (在会话间方面，我们还评估了关系GAT处理不同类型的社会信息和TIE模块的必要性。)
- The performance of DREAM-GAT is slightly poorer than ours, indicating relational-GAT can sharply catch the difference between the two social information. (DREAM-GAT的表现略差于我们的，这表明关系GAT能够敏锐地捕捉到两种社会信息之间的差异。)
- The results show that TIE module improves the performance compared with DREAM-TGRU. (结果表明，与DREAM-TGRU相比，TIE模块提高了性能。)
- This fully verifies the temporal effects in capturing users’ interests. More importantly, the more session information doesn’t mean better performance. (这充分验证了捕获用户兴趣的时间效应。更重要的是，会话信息越多并不意味着性能越好。)
DREAM utilizing 2 session’s information gets better performance than DREAM-3. Moreover, the more session information used, the longger running time it takes to run. (DREAM利用2个会话的信息比DREAM-3获得更好的性能。此外，使用的会话信息越多，运行所需的时间就越长。)

5 CONCLUSION

(1) In this paper, we propose DREAM for social RS.
(2) DREAM tries to model both users’ dynamic interests and their friends’ temporal influences. (DREAM试图模拟用户的动态兴趣和他们朋友的时间的影响。)
- Specifically, in each session, to solve the sparsity of social relations, (具体来说，在每个session中，为了解决社会关系的稀疏性，)
  - we design a GloVe-based method to increase the number of friends, （我们设计了一种基于GloVe的方法来增加朋友的数量，）
  - and utilize relational-GAT to integrate influences from friends. (利用关系GAT整合朋友的影响。)
- And then we build TIE modules to encode the outputs form historical sessions by recursively combining the features encoded by relational-GAT modules and that from last TIE module. (然后，通过递归组合关系GAT模块编码的特征和上一个TIE模块编码的特征，构建TIE模块对历史会话的输出进行编码。)
- By doing so, the user representations involve both the user’s dynamic interests and the dynamic influence from her friends. (通过这样做，用户表示既涉及用户的动态兴趣，也涉及来自其朋友的动态影响。)
(3) In the extensive experiments on the public datasets, DREAM significantly outperforms the state-of-the-art solutions. (在对公开数据集进行的大量实验中，DREAM的表现明显优于最先进的解决方案。)

REFERENCES

2020_CIKM_DREAM_ A Dynamic Relation-Aware Model for social recommendation相关推荐

2019_SIGIR_A Neural Influence Diffusion Model for Social Recommendation
[论文阅读笔记]2019_SIGIR_A Neural Influence Diffusion Model for Social Recommendation 论文下载地址: https://dl.a ...
论文笔记（SocialGCN: An Efficient Graph Convolutional Network based Model for Social Recommendation）
一个有效的基于图卷积神经网络的社交推荐模型原文链接:SocialGCN: An Efficient Graph Convolutional Network based Model for Socia ...
论文笔记（A Neural Influence Diffusion Model for Social Recommendation）
神经影响传播模型为了社交推荐原文链接:A Neural Influence Diffusion Model for Social Recommendation, SIGIR'19 原理:社交网络上应 ...
论文《A Neural Influence Diffusion Model for Social Recommendation》阅读
论文<A Neural Influence Diffusion Model for Social Recommendation>阅读论文概况 Abstract Introduction ...
Social Recommendation with Missing Not at Random Data（ICDM 2018）参考文献
第一部分带图的文章链接:https://blog.csdn.net/ciecus_csdn/article/details/84454425 第二部分算法以及实验部分链接:https://blog ...
对比关系生成模型(Comparative Relation Generative Model)
文章来源 Tkachenko M, Lauw H W. Comparative Relation Generative Model[J]. IEEE Transactions on Knowledge ...
Inf2vec: Latent Representation Model for Social Influence Embedding
Inf2vec: Latent Representation Model for Social Influence Embedding 2018 IEEE 34th International Con ...
2019_IJCAI_Deep Adversarial Social Recommendation
[论文阅读笔记]2019_IJCAI_Deep Adversarial Social Recommendation 论文下载地址: https://www.ijcai.org/Proceedings/ ...
2021_WWW_Self-Supervised Multi-Channel Hypergraph Convolutional Network for Social Recommendation
[论文阅读笔记]2021_WWW_Self-Supervised Multi-Channel Hypergraph Convolutional Network for Social Recommend ...
2021_AAAI_Knowledge-aware Coupled Graph Neural Network for Social Recommendation
[论文阅读笔记]2021_AAAI_Knowledge-aware Coupled Graph Neural Network for Social Recommendation 论文下载地址: htt ...

2020_CIKM_DREAM_ A Dynamic Relation-Aware Model for social recommendation