2021_SIGIR_ConsisRec: Enhancing GNN for Social Recommendation via Consistent Neighbor Aggregation

admin管理员组

文章数量:1623603

[论文阅读笔记]2021_SIGIR_ConsisRec: Enhancing GNN for Social Recommendation via Consistent Neighbor Aggregation

论文下载地址： https://doi/10.1145/3404835.3463028
发表期刊：SIGIR
Publish time: 2021
作者及单位:

• Liangwei Yang, Zhiwei Liu, Yingtong Dou Department of Computer Science, University of Illinois at Chicago {lyang84,zliu213,ydou5}@uic.edu
• Jing Ma Business School, Sichuan University jingma@stu.scu.edu
• Philip S. Yu Department of Computer Science, University of Illinois at Chicago psyu@uic.edu

数据集：

• Ciao and Epinions https://www.cse.msu.edu/~tangjili/datasetcode/truststudy.htm (作者在论文中公开的)
  We remove users without social links because they are out of social recommendation scope

代码：

• https://github/YangLiangwei/ConsisRec (作者在论文中公开的)

其他人写的文章

简要概括创新点：

• (1)However, most existing methods neglect the social inconsistency problem, which intuitively suggests that social links are not necessarily consistent with the rating prediction process. (然而，大多数现有方法忽略了社会不一致性问题，这直观地表明，社会联系不一定与评级预测过程一致。)
• (2)Social inconsistency can be observed from both context-level and relation-level. (社会不一致性既可以从语境层面观察，也可以从关系层面观察。)
  • Context-level: It indicates that users connected in a social graph may have discrepant item contexts. (上下文级别：它表示在社交图中连接的用户可能有不同的项目上下文)
  • Relation-level: There are multiple relations when simultaneously modeling social graph and user-item graph.
    关系级别：在同时建模社交图和用户项目图时，存在多个关系。
• (3) Therefore, we intend to empower the GNN model with the ability to tackle the social inconsistency problem. (因此，我们打算赋予GNN模型处理社会不一致问题的能力。)
  (4)we propose a novel framework to tackle the social inconsistency problem when conducting social recommendation, which is named as ConsisRec. (因此，我们提出了一个新的框架来解决在进行社会推荐时的社会不一致问题，称为ConsisRec。)
  • It is built upon a GNN model [13], aggregating neighbors to learn node embeddings. (它建立在GNN模型[13]的基础上，聚集邻居来学习节点嵌入。为了缓解社会矛盾问题，)
  • To alleviate the social inconsistency problem,
  • ConsisRec first generates a query embedding for selecting consistent neighbors. (ConsisRec首先生成一个用于选择一致邻居的查询嵌入。)
  • Then, it employs a neighbor sampling strategy for the selection process, where the sampling probability is based on our proposed consistency scores between the query embedding and neighbor embeddings. (然后，它在选择过程中采用了邻居抽样策略，抽样概率基于我们提出的查询嵌入和邻居嵌入之间的一致性得分。)
  • After sampling, it adopts relation attention to tackle the relation-level inconsistency. As such, neighbors with consistent relations are assigned with high importance factors for aggregation. Therefore, the learned node embeddings for rating prediction are aggregated from consistent contexts and relations. (抽样后，采用关系注意来解决关系水平不一致问题。因此，具有一致关系的邻居被赋予高度重要的聚集因子。因此，从一致的上下文和关系中聚合用于评级预测的学习节点嵌入。)

ABSTRACT

• (1) Social recommendation aims to fuse social links with user-item interactions to alleviate the cold-start problem for rating prediction. Recent developments of Graph Neural Networks (GNNs) motivate endeavors to design GNN-based social recommendation frameworks to aggregate both social and user-item interaction information simultaneously. (社交推荐旨在将社交链接与用户项目交互融合，以缓解评分预测的冷启动问题。图形神经网络（GNN）的最新发展促使人们努力设计基于GNN的社会推荐框架，以同时聚合社会和用户项目交互信息。)
• (2)However, most existing methods neglect the social inconsistency problem, which intuitively suggests that social links are not necessarily consistent with the rating prediction process. (然而，大多数现有方法忽略了社会不一致性问题，这直观地表明，社会联系不一定与评级预测过程一致。)
• (3)Social inconsistency can be observed from both context-level and relation-level. (社会不一致性既可以从语境层面观察，也可以从关系层面观察。)
• (4) Therefore, we intend to empower the GNN model with the ability to tackle the social inconsistency problem. (因此，我们打算赋予GNN模型处理社会不一致问题的能力。)
  Relation-level: There are multiple relations when simultaneously modeling social graph and user-item graph.
• (5)We propose to sample consistent neighbors by relating sampling probability with consistency scores between neighbors.(我们提出通过将抽样概率与邻居之间的一致性得分联系起来，对一致性邻居进行抽样。)
• (6) Besides, we employ the relation attention mechanism to assign consistent relations with high importance factors for aggregation. (此外，我们还利用关系注意机制来分配具有高度重要因素的一致关系，以进行聚合。)
• Experiments on two real-world datasets verify the model effectiveness.

CCS CONCEPTS

• Computing methodologies → Neural networks; • Information systems → Social recommendation.

KEYWORDS

Recommender System; Social Recommendation; Graph Neural Network

1 INTRODUCTION

• (1) A recommender system predicts how likely a user is interested in an item [19–21, 26, 27]. However, due to the high cost of data collection, most existing recommender systems suffer from the cold-start problem [18]. To alleviate it, we can incorporate the social information among users [15, 32, 35], which functions as a side information of the user-item interaction. Previous endeavour [2] shows that users’ online behaviors are greatly influenced by their social networks, such as the friendship on Wechat [36], following links on Twitter [16] and trusting links on shopping website [7]. Therefore, fusing the social links with user-item interactions is advantageous to improve the recommendation performance, which is defined as the social recommendation problem. (推荐系统可以预测用户对某个项目感兴趣的可能性[19-21,26,27]。然而，由于数据收集成本高，大多数现有的推荐系统都存在冷启动问题[18]。为了缓解这种情况，我们可以在用户之间加入社交信息[15,32,35]，这是用户项目交互的一个辅助信息。之前的研究[2]表明，用户的在线行为受其社交网络的影响很大，例如微信上的友谊[36]、Twitter上的关注链接[16]和购物网站上的信任链接[7]。因此，将社交链接与用户项交互融合有利于提高推荐性能，这被定义为社交推荐问题。)

• (2) The recent developments of Graph Neural Networks (GNNs) [17,19] help handle social recommendation tasks by simultaneously aggregating the information from both social graph and user-item graph [9, 24]. (图神经网络（GNNs）[17,19]的最新发展通过同时聚合来自社交图和用户项目图[9,24]的信息，帮助处理社交推荐任务。)

  • Based on the assumption that neighbors share similar contexts, (基于邻居共享相似环境的假设)
  • GNN learns node embeddings by aggregating neighbor information recursively on graphs [5]. (GNN通过在图上递归聚合邻居信息来学习节点嵌入[5])
  • SocialGCN [33, 34] proposes to enhance user embedding by simulating how users are influenced by the recursive social diffusion process. (SocialGCN[33,34]提出通过模拟用户如何受到递归社会扩散过程的影响来增强用户嵌入)
  • GraphRec [6] and GraphRec+ [7] jointly model three types of aggregation upon social graph, user-item graph and item-item graph, to learn user&item embeddings comprehensively. (GraphRec[6]和GraphRec+[7]在社交图、用户项图和项目项图上联合建模了三种类型的聚合，以全面了解用户和项目嵌入。)
  • DSCF [8] includes high-order social links through a sequential learning on random walks. (DSCF[8]通过随机游动的顺序学习包括高阶社交链接。)
  • MGNN [37] builds mutual social embedding layers to aggregate information from user-item rating graph and social graph. (MGNN[37]构建了相互的社交嵌入层，从用户项目评分图和社交图中聚合信息。)

• (2) However, most existing GNN-based social recommendation models ignore the social inconsistency problem [17]. (然而，大多数现有的基于GNN的社会推荐模型忽略了社会不一致性问题)

  • Specifically, the social inconsistency suggests that social links are not necessarily consistent with the rating prediction process. Aggregating the information from inconsistent social neighbors spoils the ability of a GNN to characterize beneficial information for recommendation. (具体而言，社会不一致性表明，社会联系不一定与评级预测过程一致。将来自不一致的社会邻居的信息聚合在一起，会破坏GNN描述有益于推荐的信息的能力。)

• (3) The social inconsistency can be categorized into two levels: (社会不一致可分为两个层面：)

  • Context-level: It indicates that users connected in a social graph may have discrepant item contexts. We demonstrate the context-level social inconsistency in Figure 1(a). We use dash lines and the solid lines to represent user-item ratings and social connections, respectively. As seen, u 3 u_3 u3​ would be u 2 u_2 u2​’s inconsistent neighbor because the items of u s u_s us​ are all books, while u 2 u_2 u2​’s rated items all belongs to sports. They have rather discrepant item contexts. (上下文级别：它表示在社交图中连接的用户可能有不同的项目上下文。我们在图1（a）中展示了语境层面的社会不一致性。我们使用虚线和实线分别表示用户项目评分和社会关系。如图所示， u 3 u_3 u3​是 u 2 u_2 u2​的不一致性邻居，因为 u 3 u_3 u3​的项目​都是书，而 u 2 u_2 u2​项目都属于体育项目。它们有相当不一致的项目上下文。)
  • Relation-level: There are multiple relations when simultaneously modeling social graph and user-item graph. For example, besides social relations, we also distinguish user-item relations by their rating values. In Figure 1(a), we observe the u 1 u_1 u1​ and u 2 u_2 u2​ are social neighbors and both connected with t 1 t_1 t1​. However, u 1 u_1 u1​ highly likes t 1 t_1 t1​(5 score) while u 2 u_2 u2​ dislikes it (1 score). It leads to the relation-level inconsistency because though socially connected, they are of inconsistent item preference. (关系级别：在同时建模社交图和用户项目图时，存在多个关系。 例如，除了社会关系，我们还通过用户项目的评分值来区分用户项目关系。在图1（a）中，我们观察到了 u 1 u_1 u1​和2 u 2 u_2 u2​是社会邻居，两者都与 t 1 t_1 t1​有关​. 然而 u 1 u_1 u1​非常喜欢 t 1 t_1 t1​（5分）而 u 2 u_2 u2​不喜欢（1分）。这导致了关系层面的不一致，因为尽管他们有社会联系，但他们的项目偏好不一致。)

• (4) To this end, we intend to empower the GNN model to solve the social inconsistency problem, which is non-trivial. (为此，我们打算授权GNN模型来解决社会不一致性问题，这是非常重要的。)

  • On the one hand, the contexts for both users and items are rather complex and difficult to express explicitly. (一方面，用户和项目的上下文都相当复杂，很难明确表达。)
  • On the other hand, we should model multiple relations simultaneously and distinguish the consistent neighbors. (另一方面，我们应该同时建模多个关系，并区分一致的邻居。)

• (5) Therefore, we propose a novel framework to tackle the social inconsistency problem when conducting social recommendation, which is named as ConsisRec. (因此，我们提出了一个新的框架来解决在进行社会推荐时的社会不一致问题，称为ConsisRec。)

  • It is built upon a GNN model [13], aggregating neighbors to learn node embeddings. (它建立在GNN模型[13]的基础上，聚集邻居来学习节点嵌入。为了缓解社会矛盾问题，)
  • To alleviate the social inconsistency problem,
  • ConsisRec first generates a query embedding for selecting consistent neighbors. (ConsisRec首先生成一个用于选择一致邻居的查询嵌入。)
  • Then, it employs a neighbor sampling strategy for the selection process, where the sampling probability is based on our proposed consistency scores between the query embedding and neighbor embeddings. (然后，它在选择过程中采用了邻居抽样策略，抽样概率基于我们提出的查询嵌入和邻居嵌入之间的一致性得分。)
  • After sampling, it adopts relation attention to tackle the relation-level inconsistency. As such, neighbors with consistent relations are assigned with high importance factors for aggregation. Therefore, the learned node embeddings for rating prediction are aggregated from consistent contexts and relations. (抽样后，采用关系注意来解决关系水平不一致问题。因此，具有一致关系的邻居被赋予高度重要的聚集因子。因此，从一致的上下文和关系中聚合用于评级预测的学习节点嵌入。)

• (6) The code is available online at https://github/YangLiangwei/ConsisRec.

• (7) The contributions of this paper are listed as follows:

  • To the best of our knowledge, we are the first work empowering the GNN model to tackle the social inconsistency problem when conducting social recommendation. (据我们所知，我们是第一个赋权GNN模型在进行社会推荐时解决社会不一致问题的工作。)
  • We propose a novel framework, ConsisRec, to learn consistent node embeddings for rating prediction. （我们提出了一个新的框架ConsistRec，用于学习一致的节点嵌入以进行评级预测。）
  • Experiments on two real-world datasets show the effectiveness of ConsisRec. Detailed analyses of ConsisRec justify its efficacy.

2 PRELIMINARIES

• (1) Social recommendation problem consists of two set of entities,

  • a user set U = { u 1 , u 2 , . . . , u m } \mathcal{U} = \{u_1, u_2, ..., u_m\} U={u1​,u2​,...,um​}
  • and an item set T = { t 1 , t 2 , . . . , t n } \mathcal{T} = \{t_1, t_2, ..., t_n\} T={t1​,t2​,...,tn​}, where m m m and n n n are the total number of users and items, respectively.
  • It includes two types of information,
  • an incomplete rating matrix R ∈ R m × n R \in R^{m \times n} R∈Rm×n
  • and the user-user social graph G s = { u , E S } \mathcal{G}_s = \{\mathcal{u}, \mathcal{E}_S\} Gs​={u,ES​}.
  • The rating R u , t R_{u, t} Ru,t​ denotes user u u u’s preference to item t t t, where higher score is interpreted as more preferring.
  • An social edge ( u i , u j ) ∈ E S (u_i, u_j) \in \mathcal{E}_S (ui​,uj​)∈ES​ indicates that user i i i has an social connection with j j j, e.g., trust.

• (2) The objective of social recommendation is to complete the rating matrix by fusing the rating matrix and the social graph. Therefore, we solve social recommendation problem by constructing a heterogeneous graph G = V , E r ∣ r = 1 R \mathcal{G} = {\mathcal{V}, {\mathcal{E}_r|}^R_{r=1}} G=V,Er​∣r=1R​,

  • where V \mathcal{V} V denotes both user and item nodes
  • and E S \mathcal{E}_S ES​ denotes edges upon relation r r r.
  • Besides user-user and item-item links, we also distinguish the user-item links by their rating scores. The score set varies on different datasets.
  • E.g., Ciao dataset [30] has 6 rating values, i.e., {0,1,2,3,4,5}. Hence, the edges on Ciao have 8 types, i.e., R R R = 8, one being social relation, one being item-item relation and the others being different rating values. (例如，Ciao数据集[30]有6个评级值，即{0,1,2,3,4,5}。因此，Ciao上的边缘有8种类型，即 R R R=8，一种是社会关系，一种是项目关系，另一种是不同的评级值。)

3 PROPOSED MODEL

The framework of ConsisRec is shown in Figure 1(b)1. It has embedding layer, query layer, neighbor sampling and relation attention.

3.1 Embedding Layer

• Following existing works [10, 33], we maintain an embedding layer E ∈ R d × ( m + n ) E \in R^{d \times(m+n)} E∈Rd×(m+n),
  • each column of which represents the trainable embedding for each node. (其中的每一列表示每个节点的可训练嵌入。)
  • We can index it to retrieve the embedding of a node v ∈ U ∪ T v \in \mathcal{U} \cup \mathcal{T} v∈U∪T as e v ∈ R d e_v \in R^d ev​∈Rd. (我们可以索引它来检索节点的嵌入)
  • In the following sections, without specific statements, we use an index v v v to denote a node, which can either be a user or an item, while u u u and t t t specifically denote a user and an item node. (在以下部分中，在没有特定语句的情况下，我们使用索引 v v v来表示节点，它可以是用户或项目，而 u u u和 t t t则具体表示用户和项目节点。)
  • Apart from node embeddings, we also train a relation embedding vector for each relation r r r to characterize relation-level social inconsistency, denoted as e r e_r er​. (除了节点嵌入外，我们还为每个关系 r r r训练一个关系嵌入向量来描述关系级的社会不一致性，即 e r e_r er​.)

3.2 Query Layer

• (1) To overcome the social inconsistency problem, we should aggregate consistent neighbors to learn node embeddings. Since social inconsistency are both in context-level and relation-level, we should distinguish consistent neighbors for each pair ( u , t ) (u, t) (u,t). Therefore, ConsisRec employs a query layer to exclusively select consistent neighbors for the query pair ( u , t ) (u, t) (u,t). It generates a query embedding by mapping the concatenation of user and item embeddings: (为了克服社会不一致性问题，我们应该聚集一致的邻居来学习节点嵌入。由于社会不一致性既存在于语境层面，也存在于关系层面，我们应该区分每一对 ( u , t ) (u, t) (u,t)的一致邻居。 因此，ConsisRec使用一个查询层专门为查询对 ( u , t ) (u, t) (u,t)选择一致的邻居。它通过映射用户和项目嵌入的串联来生成查询嵌入：)
  
  • where q u , t q_{u, t} qu,t​ is the query embedding ,
  • e u , e t ∈ R d e_u, e_t \in R^d eu​,et​∈Rd are the embedding for node u u u and t t t, respectively,
  • ⊕ \oplus ⊕ denotes concatenation,
  • W q ∈ R 2 d × d W_q \in R^{2d \times d} Wq​∈R2d×d is the mapping matrix,
  • and σ \sigma σ is a ReLU activation function.
• (2) We design a query layer to dynamically sample neighbors based on different items. (我们设计了一个查询层，根据不同的项目动态地对邻居进行采样)
  • It is because when users buy different items, they would inquire different friends.
  • Thus, u u u’s rating score of t t t is related to friends who are familiar with this query item t t t. (因此， u u u对 t t t的评分与熟悉该查询项 t t t的朋友有关)

3.3 Neighbor Sampling

• (1) Neighbor sampling has been applied to GNN to boost training [3, 4, 38] and improve ranking performance [25]. Compared with previous work, ConsisRec aims to deal with the inconsistency problem in social recommendation, and dynamic samples different social neighbors based on different items. (近邻抽样已应用于GNN，以增强训练[3,4,38]，并提高排名性能[25]。与以往的工作相比，ConsistRec旨在解决社会推荐中的不一致问题，并基于不同的项目对不同的社会邻居进行动态采样。)

• (2) Next, we present how to sample neighbors for learning the embedding of u u u and t t t. The framework of ConsisRec to aggregate node embeddings can be formalized as: (接下来，我们将介绍如何采样邻居以学习uu和tt的嵌入。ConsistRec到聚合节点嵌入的框架可以形式化为：)
  

  • where σ \sigma σ is a ReLU activation function,
  • h v ( l ) ∈ R d h^{(l)}_v \in R^d hv(l)​∈Rd is the hidden embedding of node v v v at l l l-th layer,
  • N v N_v Nv​ is the sampled neighbors of node v v v,
  • AGG is an aggregation function,
  • and W ( l ) ∈ R 2 d × d W^{(l)} \in R^{2d \times d} W(l)∈R2d×d is the mapping function.
  • h v ( 0 ) h^{(0)}_v hv(0)​ is the initial node embedding of v v v, i.e., e v e_v ev​.

• (3) Instead of equally aggregating all neighbors, we should emphasize more on consistent neighbors while ignoring those inconsistent neighbors. (我们应该更多地强调一致的邻居，而忽略那些不一致的邻居，而不是平等地聚合所有邻居。)

  • Therefore, we propose to use neighbor sampling method to select those consistent neighbors. The sampling probability for neighbor node i i i at l l l-th layer is defined by the consistency score between query q q q and all the neighbors as: (因此，我们提出使用邻居抽样方法来选择那些一致的邻居。 l l l-th层邻居节点 i i i的抽样概率由查询 q q q与所有邻居之间的一致性得分定义为：)
    
  • where s ( l ) ( i ; q ) s^{(l)}(i; q) s(l)(i;q) denotes the consistency score between the neighbor i i i and the query q q q in l l l-th GNN layer. It is defined as: (表示在 l l l-th GNN层中邻居 i i i和查询 q q q之间的一致性得分。它被定义为)
    
    • where h i ( l ) h^{(l)}_i hi(l)​ denotes the node embedding of node i i i at l l l-th layer.
    • For both nodes u u u and t t t, during the inference of rating score, we use the same query embedding. (对于节点 u u u和 t t t，在评级分数推断过程中，我们使用相同的查询嵌入。)
    • Thus, we ignore the subscript and write it as q q q for simplicity. (因此，为了简单起见，我们忽略了下标，将其写为 q q q)
    • We present this process as the sampling blocks in Figure 1(b),
      • where the probabilities for neighbors are denoted as p i p_i pi​.
    • The number of sampled neighbors is proportional to the total number of neighbors, where the ratio is 0 ≤ γ ≤ 1 0 ≤ \gamma ≤ 1 0≤γ≤1. As such, we sample more neighbors if a node is connected to more nodes. (采样的邻居数与邻居总数成正比，其中比率为0≤ γ \gamma γ≤1.因此，如果一个节点连接到更多节点，我们会对更多邻居进行采样。)

3.4 Relation Attention

• (1) After sampling the neighbors, we should aggregate their embeddings as illustrated in Eq. (2). However, the relation-level social inconsistency suggests that we should distinguish different relations. (在对邻域进行采样后，我们应该将它们的嵌入进行聚合，如公式（2）所示。然而，关系层面的社会不一致性表明，我们应该区分不同的关系。)
  • To this end, we apply a relation attention module in ConsisRec for those sampled neighbors. It learns the importance of those sampled nodes by considering the associated relations. (为此，我们在ConsistRec中为这些采样的邻居应用了关系注意模块。它通过考虑关联关系来学习这些采样节点的重要性。)
• (2) The relation attention assigns an importance factor α i \alpha_i αi​ for each sampled node i i i. We can rewrite the AGG function in Eq. (2) as:
  
  • where α i ( l ) \alpha^{(l)}_i αi(l)​ is the importance of the i i i-th neighbor sampled from Eq. (3)
  • and Q Q Q denotes the total number of sampled neighbors.
  • Assuming the relation for the edge ( v , i ) (v, i) (v,i) is r i r_i ri​, we calculate α i \alpha_i αi​ by adopting the self-attention mechanism as:
    
    • where e r i ∈ R d e_{r_i} \in R^d eri​​∈Rd represents the relation embedding of relation r i r_i ri​ and w a ∈ R 2 d w_a \in R^{2d} wa​∈R2d is trainable parameter for the self-attention layer and α i \alpha_i αi​ is the attention weights. We illustrate the relation attention as the green block in Figure 1(b).

3.5 Rating Prediction and Optimization

• (1) After L L L layer propagation, we obtain the embedding of u u u and t t t, which are denoted as h u ( L ) h^{(L)}_u hu(L)​ and h t ( L ) h^{(L)}_t ht(L)​. We calculate the rating score of the user-item pair ( u , t , ) (u, t,) (u,t,) by the inner-product of embeddings as:
  
• (2) Then the loss function is defined as the Root Mean Squared Error (RMSE) between R ^ u , t \hat{R}_{u, t} R^u,t​ and ground truth rating score R u , t R_{u, t} Ru,t​ among all ( u , t ) (u, t) (u,t) pairs in E r a t i n g \mathcal{E}_{rating} Erating​, which is calculated as
  
  • where E r a t i n g \mathcal{E}_{rating} Erating​ is the set of all rating edges. We use Adam [12] as the optimizer with a weight decay rate of 0.0001 to avoid over-fitting.

4 EXPERIMENTS

4.1 Experimental Setup

4.1.1 Datasets.

• (1) Ciao and Epinions2 are two representative datasets [28–31] for studying social recommendation problem. We remove users without social links because they are out of social recommendation scope. (我们删除没有社交链接的用户，因为他们不在社交推荐范围内。)
  • Ciao has 7,317 users, 104,975 items with 111,781 social links.
  • Epinions has 18,069 users, 261,246 items with 355,530 social links.
  • We also linked items that share more than 50% of their neighbors. (我们还链接了共享50%以上邻居的项目)

4.1.2 Baselines.

• To justify the effectiveness of ConsisRec, we compare ConsisRec with 6 baseline methods, including
  • matrix factorization methods,
  • non-GNN graph embedding methods,
  • and GNN-based methods. (非GNN图嵌入方法，)
  • SoRec [22], SocialMF [11] and SoReg [23] incorporate social links with matrix factorization methods. (SoRec[22]、SocialMF[11]和SoReg[23]将社会联系与矩阵分解方法结合起来)
  • CUNE [39] adopts collaborative graph embedding methods. (CUNE[39]采用协同图嵌入方法。)
  • GCMC+SN [1] and GraphRec [6] employ GNNs for learning node embeddings. （GCMC+SN[1]和GraphRec[6]使用GNN来学习节点嵌入）

4.1.3 Evaluation Metrics.

• To evaluate the quality of the social recommendation, two common metrics,
  • Mean Absolute Error (MAE) and
  • Root Mean Square Error (RMSE), are adopted for the rating prediction task [6].
  • Note that lower values of both indicate better performance.
  • And a small improvement in both may have a significant impact on the quality of top-N recommendation [14].

4.1.4 Experimental Settings.

• Each dataset is randomly split to 60%, 20%, and 20% for the training, validation, and testing, respectively.
• The grid search is applied for hyper-parameters tuning. We searched neighbor percent in {0.2,0.4,0.6,0.8,1.0}.
• For embedding size, we search in {8,16,32,64,128,256}.
• The learning rate is searched in {0.0005,0.001,0.005,0.01,0.05,0.1}.
• The batch size is searched in {32,64,128,256,512}.
• Only one GNN layer is used for both Ciao and Epinions datasets.
• To cope with the over-fitting problem, early stopping was utilized in all experiments, i.e., stop training if the RMSE on the validation set is not improved for five epochs.

4.2 Performance Evaluation

• The experiment results of all the methods are shown in Table 1. GCMC, GraphRec, CUNE andConsisRecperform better than SoRec, SoReg and SocialMF, which shows GNN and graph embedding based methods have a better capability to aggregate neighbor information. ConsisRec achieves the best results on both Ciao and Epinions datasets. It has an 1.7% relative improvement on two datasets compared with the second-best one on average, which can be interpreted as a significant improvement [6]. The results show the benefits brought by tackling the social inconsistency problems. (所有方法的实验结果如表1所示。GCMC、GraphRec、CUNE和ConsistercPerform优于SoRec、SoReg和SocialMF，这表明基于GNN和图嵌入的方法具有更好的聚集邻居信息的能力。ConsisRec在Ciao和Epinions数据集上都能获得最佳结果。与平均排名第二的数据集相比，这两个数据集的相对改善率为1.7%，这可以解释为显著的改善[6]。结果显示了解决社会不一致问题所带来的好处。)

4.3 Ablation Study

• (1) An ablation study is further made to evaluate different components in ConsisRec. We create three variants of ConsisRec, which are A, B, and C. (进一步进行烧蚀研究，以评估ConsistRec中的不同成分。我们创建了ConsistRec的三个变体，分别是A、B和C。)

  • A is built by removing the query layer, which directly uses user embedding instead of query embedding to select the corresponding neighbors. (通过移除查询层，直接使用用户嵌入而不是查询嵌入来选择相应的邻居，从而构建了一个新的查询层。)
  • B is built by removing neighbor sampling, which aggregates all neighbors. (B是通过移除邻域采样来构建的，它聚集了所有邻域。)
  • C is built by removing relation attention, which assigns equal weights to edges with different relations. The experimental results are illustrated in Figure 2. (C是通过消除关系注意来建立的，它为具有不同关系的边赋予相等的权重。实验结果如图2所示。)

• (2) We can observe that ConsisRec consistently achieves the best performance against other variants, demonstrating that all components are necessary to yield the best results. Additionally, we observe that the variant B B B (removing neighbor sampling module) dramatically spoils the performance, which justifies the importance of selecting consistent neighbors. The worse performance of variant A and C compared with ConsisRec also proves the importance of query layer and relation attention, respectively. (我们可以观察到，与其他变体相比，ConsistRec始终实现最佳性能，这表明所有组件都是产生最佳结果所必需的。此外，我们观察到变体B（移除邻居采样模块）显著破坏了性能，这证明了选择一致邻居的重要性。与ConsistRec相比，变体A和C的性能更差，这也分别证明了查询层和关系注意的重要性。)

4.4 Parameter Sensitivity

• Influential hyper-parameters in ConsisRec includes
  • neighbor percent,
  • embedding size
  • and the learning rate.
• Due to space limitation, we only represent results on Ciao dataset in Figure 3.
  • For neighbor percent, we observe an obvious error increment when neighbor percent rising from 0.8 to 1.0, which results from aggregating those inconsistent neighbors.
  • The best embedding size on Ciao is 16. Smaller embedding size is insufficient for representing node information, while large embedding size would lead to the over-fitting problem.
  • Learning rate has a critical impact on model performance, which needs to be tuned carefully.
    

5 CONCLUSION AND FUTURE WORK

• In this paper, we identify the social inconsistency problem related to social recommendation.
• The proposed ConsisRec contains three modifications on GNN to tackle the social inconsistency problem.
• Experiment results on two real-world datasets show the effectiveness of ConsisRec.
• Future work includes better ways to filter informative neighbors and identify the inconsistency problems inherited in other graph related research directions. (未来的工作包括更好的方法来过滤信息丰富的邻居，并识别在其他与图形相关的研究方向中继承的不一致性问题。)

6 ACKNOWLEDGEMENTS

REFERENCES

本文标签： EnhancingGNNSIGIRConsisRecsocialNeighbor