2021_AAAI_Who You Would Like to Share With? A Study of Share Recommendation in Social E-commerce

[论文阅读笔记]2021_AAAI_Who You Would Like to Share With? A Study of Share Recommendation in Social E-commerce

论文下载地址： https://www.aaai.org/AAAI21Papers/AAAI-1214.JiH.pdf
发表期刊：AAAI
Publish time: 2021
作者及单位:

• Houye Ji1, Junxiong Zhu2, Xiao Wang1, Chuan Shi1*, Bai Wang1, Xiaoye Tan2, Yanghua Li2, Shaojian He2
• 1Beijing University of Posts and Telecommunications, Beijing, China
• 2Alibaba Group, Hangzhou, China
• {jhy1993,xiaowang,shichuan,wangbai}@bupt.edu.cn,{xike.zjx,yichen.lyh}@taobao.com,{xiaoye.txy,shaojian.he}@alibaba-inc.com

数据集：

• We collect data from Taobao platform, ranging from 2019/10/09 to 2019/10/14, and construct an attributed heterogeneous graph (shown in Figure 2)

代码：

其他人写的文章

简要概括创新点： (文字介绍得比较清晰，但其实不是那么好懂，细节太多了)

• (1) In this paper, we first study the problem of share recommendation in social e-commerce, whose goal is to predict whether a user will share an item to his friend. (在本文中，我们首先研究了社交电子商务中的共享推荐问题，其目标是预测用户是否会将一个项目共享给他的朋友。)
• (2) We first construct an attributed heterogeneous graph to represent share scenario (我们首先构造一个属性异构图来表示共享场景)
• (3) and propose a novel heterogeneous GNN based share recommendation model, called HGSRec. (提出了一种新的基于异构GNN的共享推荐模型HGSRec)
  • With the help of feature embedding and semantic aggregation, the proposed HGSRec learns multiple embeddings of $u,i,v$ under different meta-paths via tripartite heterogeneous GNNs, (在特征嵌入和语义聚合的帮助下，HGSRec通过三个异构GNN学习不同元路径下$u,i,v$的多个嵌入)
  • and then dynamically fuses them via dual co-attention mechanism, (然后通过双重共同注意机制将它们动态融合)
  • followed by a transitive triplet representation to model the asymmetric share action. (然后是一个传递的三重态表示来模拟不对称的共享行为。)

Abstract

• (1) The prosperous development of social e-commerce has spawned diverse recommendation demands, and accompanied a new recommendation paradigm, share recommendation. Significantly different from traditional binary recommendations (e.g., item recommendation and friend recommendation), share recommendation models ternary interactions among $<User,Item,Friend>$, which aims to recommend a most likely friend to a user who would like to share a specific item, progressively becoming an indispensable service in social e-commerce. (社交电子商务的蓬勃发展催生了多样化的推荐需求，并伴随着一种新的推荐范式——共享推荐。与传统的二元推荐（例如，物品推荐和朋友推荐）显著不同，共享推荐模型是$<User，Item，Friend>$之间的三元交互，旨在向希望共享特定物品的用户推荐最有可能的朋友，逐渐成为社会电子商务中不可或缺的服务。)
• (2) Seamlessly integrating the social relations and purchase behaviours, share recommendation improves user stickiness and monetizes the user influence, meanwhile encountering three unique challenges: (共享推荐将社交关系和购买行为无缝地结合在一起，提高了用户粘性，并将用户影响力货币化，同时面临三个独特的挑战：)
  • rich heterogeneous information, (丰富的异构信息)
  • complex ternary interaction, (复杂的三元相互作用)
  • and asymmetric share action. (不对称的分享行为)
• (3) In this paper, we first study the share recommendation problem and propose a heterogeneous graph neural network based share recommendation model, called HGSRec. (本文首先研究了共享推荐问题，提出了一种基于异构图神经网络的共享推荐模型HGSRec。)
  • Specifically, HGSRec delicately designs a tripartite heterogeneous GNNs to describe the multifold characteristics of users and items, and （具体来说，HGSRec精心设计了一个三元异构GNN来描述用户和项目的多种特征）
  • then dynamically fuses them via capturing potential ternary dependency with a dual co-attention mechanism, (然后通过双共同注意机制 捕捉潜在的三元依赖关系，动态地融合它们)
  • followed by a transitive triplet representation to depict the asymmetry of share action and predict whether share action happens. (然后是一个传递的三重态表示来描述分享行为的不对称性，并预测股票分享是否发生。)
• (4) Offline experiments demonstrate the superiority of the proposed HGSRec with significant improvements (11.7%-14.5%) over the state-of-the-arts, and online A/B testing on Taobao platform further demonstrates the high industrial practicability and stability of HGSRec. (离线实验证明了所提出的HGSRec的优越性，与现有技术相比有显著改进（11.7%-14.5%），淘宝平台上的在线A/B测试进一步证明了HGSRec的高度工业实用性和稳定性。)

1 Introduction

• (1) In the era of information explosion, the recommender system has become the most effective way to help users to discover what they are interested in enormous data. As two basic Internet applications, e-commerce and social network both provide the recommender service and therefore generate corresponding item recommendation and friend recommendation, respectively. Recently, with the thriving of online applications, there is a surge of social e-commerce (Gefen and Straub 2004), which integrates social network and e-commerce for better e-commerce service. Benefitting from rich social interactions, social e-commerce provides a new business paradigm which improves user stickiness and activeness and monetizes the user influence. For example, Facebook and Instagram integrate e-commerce into social media, while Amazon and Taobao leverage social interactions to improve e-commerce. (在信息爆炸的时代，推荐系统已经成为帮助用户发现他们对海量数据感兴趣的最有效的方式。作为两种基本的互联网应用，电子商务和社交网络都提供推荐服务，因此分别生成相应的商品推荐和朋友推荐。最近，随着在线应用的蓬勃发展，社交电子商务兴起（Gefen and Straub 2004），它将社交网络和电子商务结合起来，以提供更好的电子商务服务。得益于丰富的社交互动，社交电子商务提供了一种新的商业模式，提高了用户的粘性和积极性，并将用户影响力货币化。例如，Facebook和Instagram将电子商务融入社交媒体，而亚马逊和淘宝则利用社交互动来改善电子商务。)

• (2) With the development of social e-commerce, a new recommendation paradigm, share recommendation, has sprung up recently. In particular, share recommendation aims to predict whether a user will share an item with his friend. Such recommendation demand is ubiquitous in social e-commerce. Share recommendation has been a unique recommendation paradigm in social e-commerce, due to thefollowing characteristics. (随着社会电子商务的发展，一种新的推荐模式——共享推荐应运而生。特别是，共享推荐旨在预测用户是否会与朋友共享项目。这种推荐需求在社交电子商务中无处不在。由于以下特点，共享推荐已成为社交电子商务中一种独特的推荐模式。)

  • Firstly, share recommendation seamlessly integrates the benefits of social relations and item recommendation. Most users coexist in purchase network and social network, so a user well knows his purchasing items, as well as his friends. seamlessly integrating them, share recommendation not only enhances the stickiness and activeness of users but also monetizes the user influence (e.g., the attention economy and Internet celebrity economy). (首先，共享推荐将社交关系和项目推荐的好处无缝地结合在一起。大多数用户同时存在于购物网络和社交网络中，因此用户很了解自己的购物项目，也很了解自己的朋友。共享推荐将二者无缝结合，不仅增强了用户的粘性和主动性，还将用户影响力货币化（例如，注意力经济和互联网名人经济）。)
  • Secondly, share recommendation provides a reliable recommendation. Since the user knows both the recommended item and his friends, the share action of the user is trustworthy for his friends, which increases the recommendation reliability and thus facilitates purchase action. (其次，共享推荐提供了可靠的推荐。由于用户既知道推荐的商品，也知道他的朋友，因此用户的分享行为对他的朋友来说是值得信任的，这增加了推荐的可靠性，从而促进了购买行为。)
    

• (3) The share recommendation is significantly different from traditional recommendations, such as item recommendation (Wang et al. 2019a) and friend recommendation (Wang et al. 2014). (共享推荐明显不同于传统推荐，如项目推荐（Wang et al.2019a）和朋友推荐（Wang et al.2014）。)

  • As shown in Figure 1, we can find that
    • item recommendation aims to recommend an item to a user (i.e., essentially maximize the probability $P(i2|u2)$)
    • and friend recommendation aims to recommend a friend to a user (i.e., maximize the probability $P(u4|u2)$).
    • Note that social recommendation (Ma et al. 2008) is naturally item recommendation. Significantly different from the above binary recommendations, the goal of share recommendation is to predict the ternary interactions among $<User,Item,Friend>$, i.e., whether a user will share an item with his friend, maximizing the probability $P(u3|u2,i3)$. (请注意，社会推荐（Ma et al.2008）自然是项目推荐。与上述二元推荐显著不同，共享推荐的目标是预测$<user，item，friend>$之间的三元交互，即用户是否会与朋友共享一个项目，从而最大化概率$P(u3|u2，i3)$)

• (4) Deliberately considering the characteristics of share recommendation, we need to address the following challenges for modeling share recommendation. (考虑到股票推荐的特点，我们需要为共享推荐建模解决以下挑战。)
  

• (5) Rich Heterogeneous Information. Share recommendation usually contains complex heterogeneous information, including complex interactions among users and items, as well as rich feature information of users and items. Such an example is shown in Figure 2(a). How to handle the complex interactions and utilize the diverse features simultaneously is an urgent problem that needs to be solved. （丰富的异构信息。共享推荐通常包含复杂的异构信息，包括用户和项目之间复杂的交互，以及用户和项目丰富的特征信息。如图2（a）所示。如何处理复杂的交互，同时利用不同的特征是一个迫切需要解决的问题。）

• (6) Complex ternary interaction. Different from simple binary interaction in traditional recommendations, exemplified as $<u2,i2>$ interaction in the item recommendation and $<u2,u4>$ interaction in friend recommendation in Figure 1, share recommendation faces complex ternary interaction (e.g., $<u2,i3,u3>$ in Figure 1). We need to consider the suitability of a share action, which evaluates the matching degree of three objects (e.g.,$<u2,i3,u3>$) in the share action. According to the characteristic of the recommended item, a user will recommend it to an appropriate friend, and thus how the item influence the user (or the friend) should be considered. Taking Figure 1 as an example, the user $u_2$ will share the shoes $i_3$ to his classmate $u_3$, rather than his mom $u_1$. So we need to model the ternary interaction of user, item, and friend, considering their suitability. (复杂的三元相互作用。与传统推荐中的简单二元交互不同，如图1中项目推荐中的<u2，i2>交互和朋友推荐中的<u2，u4>交互，共享推荐面临复杂的三元交互（例如，图1中的<u2，i3，u3>）。我们需要考虑共享行为的适用性，它评估共享行为中三个对象（例如; <u2，i3，u3>;）的匹配程度。根据推荐项目的特点，用户会将其推荐给合适的朋友，因此应该考虑该项目对用户（或朋友）的影响。以图1为例，用户$u_2$​ 将分享鞋子$i_3$ 致他的同学$u_3$, 而不是他妈妈​.因此，我们需要考虑用户、项目和朋友之间的三元交互，并考虑它们的适用性。

• (7) Asymmetric Share Action. The share action is asymmetric and irreversible, which means the share action may not happen if we swap the roles of the user and the friend. As shown in Figure 1, the user $u_2$ may share a women overcoat $i_1$ to his mom $u_1$, while the user $u_1$ would not share the women overcoat $i_1$ to her son $u_2$. Therefore, a desired model should consider the asymmetry of share action. (不对称分享行为。共享行为是不对称且不可逆的，这意味着如果我们交换用户和朋友的角色，共享行为可能不会发生。如图1所示，用户$u_2$可以分享一件女式大衣给他妈妈​ , 而用户$u_1$不会分享这件女大衣​给她的儿字$u_2$​. 因此，所期望的模型应该考虑共享行为的不对称性。)

• (8) In this paper, we first study the problem of share recommendation and propose a novel Heterogeneous Graph neural network based Share Recommendation model (HGSRec).

  • We model the share recommendation system as an attributed heterogeneous graph to integrate rich heterogeneous information, (我们将分享推荐系统建模为一个属性异构图，以集成丰富的异构信息，)
  • and then we design HGSRec to learn the embeddings of $u,i,v$ and predict the probability of share action $<u,i,v>$ happening.
  • Specifically, after initializing node embedding via encoding rich node features, a tripartite heterogeneous GNNs is designed to learn the embeddings of $u,i,v$, respectively, via aggregating their meta-path based neighbors, which enables HGSRec flexibly fuse different aspects of information. (具体来说，在通过编码丰富的节点特征初始化节点嵌入后，设计了一个三元异构GNN，通过聚合其基于元路径的邻居，分别学习$u、i、v$的嵌入，从而使HGSRec能够灵活地融合不同方面的信息。)
  • Furthermore, a dual co-attention mechanism is proposed to dynamically fuse the multiple embeddings of $u$ (or $v$) under different meta-paths, considering the influence of item $i$ to user $u$ (or $v$), to improve the suitability of $<u,i,v>$. (此外，考虑到项目i对用户u（或v）的影响，提出了一种双共同注意机制，动态融合不同元路径下u（或v）的多个嵌入，以提高<u，i，v>的适用性)
  • Finally, a transitive triplet representation of $<u,i,v>$ is employed to predict whether share action happens.

• (9) The contributions of our work are summarized as follows:

  • We study a newly emerging recommendation problem, share recommendation, which aims to predict whether a user will recommend an item to his friend. Different from traditional binary recommendations, the share recommendation provides a ternary recommendation paradigm. (我们研究了一个新出现的推荐问题，共享推荐，其目的是预测用户是否会向朋友推荐一个项目。与传统的二元推荐不同，共享推荐提供了一种三元推荐范式。)
  • We propose a novel HGSRec for share recommendation, with the help of delicate designs, such as (我们提出了一种新颖的HGSRec共享推荐方案，借助于精致的设计，例如)
    • tripartite heterogeneous GNNs,
    • dual co-attention mechanism and
    • transitive triplet representation.
  • Extensive experiments on Taobao demonstrate the superiority of the proposed HGSRec with more than 10% performance improvement, compared to the state-of-the-arts.

2 Related Work

• (1) Item recommendation (Sarwar et al. 2001; Shi et al. 2018; Hu et al. 2018), which aims to predict whether one user will buy or view one item, has been extensively studied and provided great economic value. Several works (Ma et al. 2008; Fan et al. 2019b) leverage social information to further improve the performance of item recommendation. (商品推荐（Sarwar等人，2001年；Shi等人，2018年；Hu等人，2018年）旨在预测一个用户是否会购买或观看一个商品，已被广泛研究并提供了巨大的经济价值。有几项工作（Ma等人，2008年；Fan等人，2019b）利用社会信息来进一步提高项目推荐的性能。)
  • People recommendation (Kutty, Nayak, and Chen 2014; Ricci,Rokach, and Shapira 2011) aims to predict whether one user will interact with another user, such as user-friend in friend recommendation, employer-employee in job recommendation, and male-female in dating recommendation. (人员推荐（Kutty、Nayak和Chen 2014；Ricci、Rokach和Shapira 2011）旨在预测一个用户是否会与另一个用户互动，例如朋友推荐中的用户朋友、工作推荐中的雇主-雇员以及约会推荐中的男性-女性。)
  • Essentially, both item recommendation and people recommendation consider the binary interaction, such as $<User,Item>$ and $<User,User>$.
  • Significantly different from the above recommendations, share recommendation focus on ternary interaction $<User,Item,Friend>$ and predict whether one user will share one item with his friend.
• (2) Graph neural networks (Kipf and Welling 2017; V eliˇ ckovi´ c et al. 2018; Hamilton, Ying, and Leskovec 2017) generalize deep learning to graph-structured data, which usually follows the message-passing framework to receive messages from neighbors and apply neural network to update node embedding. Kipf et al. (Kipf and Welling 2017) propose graph convolutional network for node lassification, and (Hamilton, Ying, and Leskovec 2017; Veliˇ ckovi´ c et al. 2018) propose diverse aggregating functions. Several works (Ying et al. 2018; Fan et al. 2019b; Wang et al. 2019a; Wu et al. 2019) generalize GNNs to perform item recommendation. Recently, some works (Wang et al. 2019b; Fan et al. 2019a; Hu et al. 2019; Wang et al. 2020) extend GNNs for heterogeneous graph and generalize them (Fan et al. 2019a; Zhao et al. 2019) for recommendation. However, all previous GNN based recommendation methods focus on binary interaction and cannot be applied to model ternary interaction $<User,Item,Friend>$ in share recommendation.

3 Preliminaries

3.1 Definition 1. Attributed Heterogeneous Graph. (属性异质图)

• An attributed heterogeneous graph, denoted as $\mathcal{G}=(\mathcal{V},\mathcal{E},X)$,
  • where $\mathcal{V}={\mathcal{V}}_U\cup {\mathcal{V}}_I$ is the node sets,
  • $\mathcal{E}={\mathcal{E}}_S\cup {\mathcal{E}}_O$ is the edge sets,
  • $X\in R^{|\mathcal{V}|\times K}$ is an attribute matrix of nodes.
  • Here ${\mathcal{V}}_U$ and ${\mathcal{V}}_{\mathcal{I}}$ are the sets of users and items, respectively.
  • ${\mathcal{E}}_S=<{\mathcal{V}}_U,{\mathcal{V}}_U>$ denotes $User-Userinteraction$
  • and ${\mathcal{E}}_O=<{\mathcal{V}}_U,{\mathcal{V}}_I>$ denotes $User-Iteminteraction$.
  • For $u,v\in {\mathcal{V}}_U$, $v$ is $u$’s friend if $<u,v>\in {\mathcal{E}}_S$
  • and the friend set of $u$ is F ( u ) = { v ∣ < u , v > ∈ E S } \mathcal{F}(u) = \{v|<u, v> \in \mathcal{E}_S\} F(u)={v∣<u,v>∈ES​}.

Example.

• (1) Figure 2(a) shows the attributed heterogeneous graph of share recommendation. Here $u_2$ has two friends denoting as F ( u 2 ) = { u 1 , u 3 } \mathcal{F}(u_2) = \{u_1, u_3\} F(u2​)={u1​,u3​}.
• (2) Meta-path (Sunet al. 2011), a composite relation connecting two nodes, is able to extract rich semantics. As shown in Figure 2(b),
  • $User\stackrel{buy}{⟶}Item\stackrel{buy}{⟶}User$ ($U-b-I-b-U$ for short) meaning the co-buying relations,
  • $User\stackrel{social}{⟶}User$ ($U-s-U$ for short) meaning the social relations,
  • $User\stackrel{buy}{⟶}Item$ ($U-b-I$ for short) meaning buy relations,
  • and $User\stackrel{view}{⟶}Item\stackrel{view}{⟶}User$ ($U-v-I-v-U$ for short) meaning the co-viewing relations.

3.2 Definition 2. Share Recommendation. (共享推荐)

• (1) Given an attributed heterogeneous graph $\mathcal{G}=(\mathcal{V},\mathcal{E},X)$ representing a share recommendation system, share recommendation aims to predict a share action $<u,i,v>$ (formulated with $<User,Item,Friend>$, or abbreviated with $<U,I,V>$)
• (2) Specifically, the purpose of share recommendation is to recommend the most likely $Friendv\in \mathcal{F}(u)$ to $Useru\in {\mathcal{V}}_U$ who would like to share the $Itemi\in {\mathcal{V}}_I(<u,i>\in {\mathcal{E}}_O)$, i.e., $argma{x}_{v}P(v|u,i)$.
• (3) The label y u , i , v ∈ { 0 , 1 } y_{u,i,v}\in \{0,1\} yu,i,v​∈{0,1} indicates whether share action happens.

Example.

• As shown in Figure 1(c), share recommendation will recommend a most likely friend, like $u_3\in \mathcal{F}(u_2)$, to a user $u_2$ who would like to share the shoes $i_3$, which essentially maximizes the probability $P(u3|u2,i3)$.

4 The Proposed Model

In this section, we present a novel Heterogeneous Graph neural network based Share Recommendation (HGSRec). The overall framework of HGSRec is shown in Figure 3.

4.1 Initialization with Feature Embedding

• Firstly, we initialize node embedding via embedding their features. Different from ID embedding, feature embedding has two-fold benefits:
  • (1) In real applications, there are numerous of newly coming nodes every day. The feature embedding effectively generates embeddings for previously unseen nodes by utilizing their features. (在实际应用中，每天都有大量新出现的节点。特征嵌入通过利用节点的特征有效地为以前看不见的节点生成嵌入。)
  • (2) The number of features is much less than the number of nodes, which significantly reduces the number of learnable parameters. (特征的数量远少于节点的数量，这大大减少了可学习参数的数量)
• (2) For the $k$-th node feature $f_k\in R^{|f_k|\ast 1}$, we initialize a feature embedding matrix $M^{f^k}\in R^{d\ast |f_k|}$,
  • where $|f_k|$ means the number of values of feature $f_k$
  • and $d$ is the dimension of feature embedding.
  • The embedding of $u$’s $k$-th feature is shown as follows:
    
• (2) Considering all the features of user $u$, we can get the initial user embedding $x_u$, as follows:
  
  • where || denotes the concatenate operation, $W_U$ and $b_U$ denote the weight matrix and bias vector, respectively. The same process can be done for item/friend embedding.

4.2 Tripartite Heterogeneous Graph Neural Networks

• (1) Here we propose tripartite heterogeneous GNNs to learn embeddings of $u,i,v$ via corresponding heterogeneous GNN (i.e., $HeteGN{N}^U,HeteGN{N}^I,andHeteGN{N}^V$), respectively.
  • Heterogeneous GNN usually follows a hierarchical manner: (异构GNN通常遵循分层方式)
    • It first aggregates information from one kind of neighbors via one meta-path and learns the semantic-specific node embeddings in node-level. (它首先通过一条元路径从一类邻居处聚集信息，并在节点级学习语义特定的节点嵌入。)
    • Then, it aggregates multiple semantics from different meta-paths and fuses a set of semantic-specific node embeddings in semantic-level. (然后，它聚合来自不同元路径的多个语义，并在语义级别融合一组语义特定的节点嵌入。)
• (2) Specifically, given one user $u$ and $k_1$ user-related meta-paths { Φ 1 U , Φ 2 U , ⋅ ⋅ ⋅ , Φ k 1 U } \{\Phi^U_1, \Phi^U_2,· · · ,\Phi^U_{k_1}\} {Φ1U​,Φ2U​,⋅⋅⋅,Φk1​U​}, $HeteGN{N}^U$ is able to get $k_1$ semantic-specific user embeddings { x u Φ 1 U , x u Φ 2 U , ⋅ ⋅ ⋅ , x u Φ k 1 U } \{x^{\Phi^U_1}_u,x^{\Phi^U_2}_u,· · · ,x^{\Phi^U_{k1}}_u \} {xuΦ1U​​,xuΦ2U​​,⋅⋅⋅,xuΦk1U​​}.(特定语义的用户嵌入)
  
• (3) Note that the number of meta-path based neighbors of different nodes could be quite different, so we need to sample fixed number of neighbors. Random sampling strategy causes heavy computation consumption and missing important nodes. Here we propose a top-N semantic sampling strategy:(前N语义采样策略)
  • (1) If the number of meta-path based neighbors is more than fixed number $N$, we sample top-N meta-path based neighbors based on connection strength (e.g., how many times a user view an item). (如果基于元路径的邻居的数量超过固定数量N，我们将根据连接强度（例如，用户查看一个项目的次数）对前N个基于元路径的邻居进行采样。)
  • (2) Or else, we adopt resample to get N meta-path based neighbors. (或者，我们采用重采样来获得N个基于元路径的邻居)
• (4) Given a user $u$ and corresponding meta-path ${\Phi }^U$, we propose a novel semantic aggregator $SemAg{g}_u^{{\Phi }^U}$ to aggregate sampled neighbors $N_u^{{\Phi }^U}$ and obtain the meta-path based embedding $x_u^{N^{{\Phi }_u^U}}$, as follows:
  
• (5) Considering the time efficiency, we adopt $MeanPooling$ to accelerate aggregating processing for faster prediction. The semantic aggregator $SemAg{g}_u^{{\Phi }^U}$ is shown as follows:
  
• (6) To emphasize the property of user $u$ explicitly, we concatenate initial embedding $x_u$ and meta-path based embedding $x_u^{N_u^{{\Phi }^U}}$ and get the semantic-specific user embedding $x_u^{{\Phi }^U}$,
  
  • where $W^{{\Phi }^U}$ and $b^{{\Phi }^U}$ denote the weight matrix and bias vector for meta-path ${\Phi }^U$, respectively.
    • Given a set of user-related meta-paths { Φ 1 U , Φ 2 U , ⋅ ⋅ ⋅ , Φ k 1 U } \{\Phi^U_1,\Phi^U_2,· · · ,\Phi^U_{k1}\} {Φ1U​,Φ2U​,⋅⋅⋅,Φk1U​}, we can get $k_1$ semantic-specific user embeddings $x_u^{{\Phi }_1^U},x_u^{{\Phi }_2^U},\cdot \cdot \cdot ,x_u^{{\Phi }_{k_1}^U}$ which describe the characteristics of user $u$ from different aspects.
    • The same process can be done via $HeteGN{N}^V$ to learn multiple semantic-specific embeddings { x v Φ 1 V , x v Φ 2 V , ⋅ ⋅ ⋅ , x v Φ k 2 V } \{x^{\Phi^V_1}_ v, x^{\Phi^V_2}_v,· · · ,x^{\Phi^V_{k_2}}_v \} {xvΦ1V​​,xvΦ2V​​,⋅⋅⋅,xvΦk2​V​​} of friend $v$.
    • Since the characteristic of the item is much simple and stable than the user, we only adopt one meta-path ${\Phi }^I$ to get the embedding $x_i^{{\Phi }^I}$ of item $i$ via $HeteGN{N}^I$.

4.3 Dual Co-Attention Mechanism

• (1) After obtaining a set of semantic-specific node embeddings (e.g., { x u Φ 1 U , x u Φ 2 U , ⋅ ⋅ ⋅ , x u Φ k 1 U } \{x^{\Phi^U_1}_u, x^{\Phi^U_2}_u,· · · ,x^{\Phi^U_{k_1}}_ u\} {xuΦ1U​​,xuΦ2U​​,⋅⋅⋅,xuΦk1​U​​}), we aim to fuse them properly based on the complex ternary interactions $<u,i,v>$.
  • So a dual co-attention mechanism is designed to dynamically fuse the embeddings of $u$ (or $v$) under different meta-paths, considering the effect of item $i$.
  • which consist of co-attention mechanism $CoAt{t}_{U,I}$ for $<U,I>$ and co-attention mechanism $CoAt{t}_{V,I}$ for $<V,I>$.
  • Specifically, it learns the interaction-specific attention values of meta-paths for $<u,i,v>$ and get the most appropriate embedding of $u$, $v$, with the following benefits:
  • (1) It reinforces the dependency of $<u,i,v>$, making HGSRec more integrated. (它增强了<u，i，v>的依赖性，使HGSRec更加集成)
  • (2) It dynamically fuses the embeddings of $u$ (or $v$), improving share suitabilities.
• (2) Taking $<U,I>$ as an example, the co-attention mechanism $CoAt{t}_{U,I}$ aims to learn a set of interaction-specific co-attention weights $w_{u,i}^{{\Phi }_1^U},w_{u,i}^{{\Phi }_2^U},\cdot \cdot \cdot ,w_{u,i}^{{\Phi }_{k_1}^U}$ for user $u$,
  
• (3) Specifically, we concatenate the semantic-specific embedding of $u$ and $i$ and project them into co-attention space. Then, we adopt a co-attention vector $q_{U,I}$ to learn the importances of meta-paths for user $u$. The importance of meta-path ${\Phi }_m^U$ for $u$ in the interaction $<u,i>$, denoted as ${\alpha }_{u,i}^{{\Phi }_m^U}$,
  
  • where $W^{U,I}$ and $b^{U,I}$ denote the weight matrix and bias vector, respectively.
• (4) After obtaining the importances of meta-paths, we normalize them via softmax to get the co-attention weight $w_{u,i}^{{\Phi }_m^U}$ of meta-path ${\Phi }_m^U$, shown as follows:
  
  • where $w_{u,i}^{{\Phi }_m^U}$ reflects the contribution of meta-path ${\Phi }_m^U$ in improving share suitability. Larger $w_{u,i}^{{\Phi }_m^U}$ means the meta-path ${\Phi }_m^U$ of $u$ is more suitable to the item $i$ which makes higher contribution in improving the suitability of $<u,i,v>$. With the learned weights as coefficients, we can obtain the fused embedding $h_u$ of $u$, shown as follows:
    
• (5) Obviously, $h_u$ is dynamically changed with regard to different co-attention weights, where the co-attention weights are dynamically changed with regard to different items. (显然根据不同的共同注意权重动态变化，其中共同注意权重根据不同的项目动态变化。)
• (6) Similar to $CoAt{t}_{U,I}$, $CoAt{t}_{V,I}$ learns a set of co-attention weights $w_{v,i}^{{\Phi }_1^V},w_{v,i}^{{\Phi }_2^V},\cdot \cdot \cdot ,w_{v,i}^{{\Phi }_{k_2}^V}$ for friend $v$ and get the fused friend embeddings $h_v$. Since we only select one meta-path for item, the fused embedding $h_i$ of item $i$ is actually $x_i^{{\Phi }^I}$.

4.4 Transitive Triplet Representation

• (1) To predict the share action $<u,i,v>$, we need to construct a triplet representation $r_{u,i,v}$ based on $h_u,h_i,h_v$. We first project all types of nodes in $<U,I,V>$ into the same space via three type-specific MLPs, shown as follows:
  
• (2) A simple way to construct the triplet representation $r_{u,i,v}$ is to concatenate all node embeddings (a.k.a., $z_u$||$z_i$||$z_v$). However, the simple concatenation cannot explicitly capture the remarkable characteristics of share action: (然而，简单的串联并不能明确地捕捉共享行为的显著特征：)
  • (1) The share recommendation actually aims to rank candidate friends based on both user and item (e.g., calculate the similarity between $z_u+z_i$ and $z_v$), so the share action is asymmetric and the roles of user and friend cannot be exchanged. (共享推荐实际上旨在根据用户和项目对候选朋友进行排名（例如，计算$z_u+z_i$和$z_v$之间的相似性 ), 因此，共享行为是不对称的，用户和朋友的角色无法交换。)
  • (2) The item describes the transition between user and friend, so it is an indispensable bridge in establishing share action. (该项目描述了用户和朋友之间的过渡，因此它是建立共享行动不可或缺的桥梁)
• (3) Inspired by relational translation (Antoine et al. 2013), we propose a transitive triplet representation $r_{u,i,v}$ to explicitly model the characteristics of share action via item-translating, shown as follows: (受关系翻译（Antoine et al.2013）的启发，我们提出了一种传递三元组表示$r_{u，i，v}$通过项目翻译对共享行为的特征进行明确建模，如下所示)
  
  • where $|\cdot |$ denotes the absolute operation.
• (4) Then, we feed $r_{u,i,v}$ into MLP and get the predict score ${\hat{y}}_{u,i,v}$, as follows:
  
  • where $W$ and $b$ denote the weight vector and bias scalar, respectively.
• (5) Finally, we calculate cross-entropy loss,
  
  • where $y_{u,i,v}$ is the label of the triplet,
  • $D$ denotes the dataset.
  • Then, we analyze the space complexity of HGSRec.
• (6) The learnable parameters in HGSRec mainly come from embedding matrixes rather than neural networks. Assuming we have $A$ node IDs and $B$ node features, the number of parameters of ID embedding and feature embedding are $O(A\ast d)$ and $O(B\ast d)$, respectively. In the share scenario, the number of IDs $A$ (usually billion-level) is significantly more than the number of features $B$ so as to $O(A\ast d)\gg O(B\ast d)$, which means HGSRec is able to efficiently handle large-scale data.

5 Experiments

5.1 Datasets

• We collect data from Taobao platform, ranging from 2019/10/09 to 2019/10/14, and construct an attributed heterogeneous graph (shown in Figure 2).
  • Each sample contains a share action $<u,i,v>$ and corresponding label y u , i , v ∈ { 0 , 1 } y_{u,i,v}\in \{0,1\} yu,i,v​∈{0,1}.
  • We select four meta-paths including $U-s-U$, $U-b-I-b-U$ and $U-v-I-v-U$ for the user
  • and $U-b-I$ for the item. (挑了4条元路径)
  • In offline experiments, we use the last day (i.e., 2019/10/14) as validation set and the previous 3/4/5 days as training sets, marked as 3-days, 4-days, and 5-days, respectively.
  • To comprehensively evaluate the results, we vary the size of each training set from 40% to 100%. The details of the datasets are shown in Table 1.
    

5.2 Baselines

• (1) We select feature based models (i.e., LR, DNN, and XGBoost) and
  • GNN models (i.e., GraphSAGE, IGC, and MEIRec) as baselines.
  • Since IGC and MEIRec cannot handle ternary recommendation, we also provide tripartite versions (i.e., IGC+ and MEIRec+) for share recommendation.
  • To validate delicate designs in HGSRec, we also test two variants of HGSRec ($HGSRe{c}_{\backslash att}$ and $HGSRe{c}_{\backslash tra}$).
  • Note that although deep models depend on randomness whose performances change with different random seeds, their performances on large-scale Taobao datasets are quite stable (a.k.a, the variance of HGSRec less than 0.001). (请注意，尽管深度模型依赖于随机性，其性能随不同的随机种子而变化，但它们在淘宝大规模数据集上的性能相当稳定（也就是说，HGSRec的方差小于0.001）。)
• LR/DNN/XGBoost: They are classical algorithms for industry. We concatenate the feature of user, item and friend as the model input and predict the share action $<u,i,v>$.
• GraphSAGE (SAGE for short) (Hamilton, Ying, and Leskovec 2017): It is a classical GNN which leverages sampler and aggregator to embed homogeneous graph. Since GraphSAGE cannot perform share recommendation directly, we ignore the influence of item and utilize it to perform people-to-people recommendation (a.k.a, $<u,v>$). (它是一种利用采样器和聚合器嵌入齐次图的经典GNN。由于GraphSAGE无法直接执行共享推荐，因此我们忽略了项目的影响，并利用它执行人与人之间的推荐)
• IGC/IGC+ (Zhao et al. 2019): It is a HeteGNN based recommendation model. Since IGC cannot perform share recommendation directly, (是一个基于HeteGNN的推荐模型。由于IGC不能直接执行股票推荐，)
  • we first recall all candidate friends of user $u$ and then predict the $<i,v>$. (我们首先召回用户u的所有候选朋友，然后预测)
  • Then, we extend IGC as IGC+ to learn the embedding of $u,i,v$ and concatenate them to predict $<u,i,v>$.
• MEIRec/MEIRec+ (Fan et al. 2019a): It is a HeteGNN based recommend model. Similar to IGC+, we also extend MEIRec as MEIRec+ for share recommendation.
• $HGSRe{c}_{\backslash att}$: It is a variant of HGSRec, which removes the dual co-attention mechanism and employs the simple average strategy on all meta-paths for recommendation.
• $HGSRe{c}_{\backslash tra}$: It is a variant of HGSRec, which removes the transitive triplet representation and concatenates the embeddings of $u,i,v$ for recommendation.

5.3 evaluation metric && optimizer && parameter settings

• We select AUC as the evaluation metric,
• RMSProp as optimizer.
• We uniformly set feature embedding to 8, node embedding to 128, batch size to 1024, learning rate to 0.01 and dropout rate to 0.6 for deep models.
• For XGBoost, we set tree depth to 6, tree number to 10.
• For LR, we set the L1 reg to 1.
• For HeteGNNs, we sample 5, 10, 2 neighbors via $U-s-U$, $U-v-I-v-U$ , $U-b-I-b-U$ to learn multiple user embeddings and sample 50 neighbors via $U-b-I$ to learn item embedding.

5.4 Performance Evaluation

As shown in Table 2, we have the following observations:

• (1) HGSRec consistently performs better than all baselines with significant improvements. Compared to the best baseline, the improvements are up to 11.7%-14.5%, indicating the superiority of HGSRec. (HGSRec的表现始终优于所有基线，并有显著改善。与最佳基线相比，改善率高达11.7%-14.5%，表明HGSRec的优越性。)
• (2) Most of GNNs (i.e., GraphSAGE, IGC, and MEIRec) outperform feature based methods (i.e., LR, DNN, and XGBoost), indicating the importance of structure information. When deeper insight into these methods, we can find, if employing ternary interactions, the tripartite versions (i.e., IGC+ and MEIRec+) significantly outperform the original versions. It further confirms the benefits of modeling ternary interaction for share recommendation. (大多数GNN（即GraphSAGE、IGC和MEIRec）都优于基于特征的方法（即LR、DNN和XGBoost），表明结构信息的重要性。深入了解这些方法后，我们可以发现，如果采用三元相互作用，三方版本（即IGC+和MEIRec+）的性能显著优于原始版本。它进一步证实了为股票推荐建模三元交互的好处。)
• (3) Comparing the performance of HGSRec with its variants, we can find HGSRec achieves the best performance.
  • The degradation of $HGSRe{c}_{\backslash att}$ indicates the effectiveness of the dual co-attention mechanism,
  • while the degradation of $HGSRe{c}_{\backslash tra}$ validates the superiority of transitive triplet representation.
  • Note that the degradation of $GSRe{c}_{\backslash tra}$ is much more significant than that of $HGSRe{c}_{\backslash att}$, which implies that transitive triple representation may make higher contribution than dual co-attention mechanism.

5.5 Attention Analysis

• (1) The dual co-attention mechanism can dynamically fuse multiple embeddings of $User$ and $Friend$ with regard to different Items and improve the share suitabilities.
  • We first present the macro-level analysis via the box-plot figure of attention distributions over $User$ on 3-day dataset in Figure 4(a).
  • Note that attention values distributions over $Friend$ also show similar phenomenons. As can be seen, the attention distribution of meta-paths are different, and the attention values of $U-b-I-b-U$ is the largest with a higher variance, which illustrates that this meta-path is the most important for most users. The reason is that $U-b-I-b-U$ is related to user purchasing behavior which reflects the strongest user preference. The higher variance of $U-b-I-b-U$ also implies its importances varies greatly for different users. We further test HGSRec with single meta-path and show their performances with the corresponding averaged attention values in Figure 4(b). Consistent with attention distribution, $U-b-I-b-U$ is the most useful meta-path which achieves the highest AUC and gets the largest attention value.

• (2) We further present a case study to show the potential interpretability of HGSRec. We select a share action $<u707,i586,v198>$, where a user $u707$ shares an eye shadow $i586$ to his friend $v198$. Note that eye shadow $i586$ belonging to the category of $Makeup/Perfume$.
  • As can be seen in Figure 5, the learned attention values between $<u707,i586i>$ and $<v198,i586>$ are significantly different from each other.
  • The interaction between $<u707,i586>$ mainly depends on $U-s-U$,
  • while the interaction between $<v198,i586>$ mainly depends on $U-s-U$ and $U-v-I-v-U$.
  • By inspecting into the dataset, we found that most of $u707$’s friends like to buy $Makeup/Perfume$ and some of them buy more than 20 items belonging to this category. It explains why the $U-s-U$ plays the key role in $<u707,i586>$.
  • The users who connect to $v198$ via $U-s-U$ and $U-v-I-v-U$ also like buying the items belonging to $Makeup/Perfume$ which explains why $U-s-U$ and $U-v-I-v-U$ both play the key roles in $<v198,i586>$.
  • In summary, the proposed HGSRec is able to learn appropriate attention values for user and friend with regard to different items in different share actions, providing potential interpretability for recommendation results. (总之，建议的HGSRec能够为用户和朋友了解关于不同分享行为中不同项目的适当注意值，为推荐结果提供潜在的可解释性。)
    

5.6 Effects of Different Meta-paths

• To further investigate the effect of different meta-paths, we test the performance of HGSRec on 3-day dataset via adding four meta-paths ($U-b-I$, $U-v-I-v-U$, $U-b-I-b-U$, and $U-s-U$) one by one.
  • As shown in Figure 6, with the addition of meta-paths, the performance of HGSRec improves consistently. It demonstrates more comprehensive information extracted by meta-paths indeed improves node embedding.
  • Note that different meta-paths have different impacts. When adding $U-b-I-b-U$, HGSRec achieves the largest improvement. Similar to the results in Figure 4(a), $U-b-I-b-U$ is the most informative path, because it directly reflects user intention.

5.7 Online Experiments

• (1) We deploy HGSRec on Taobao APP for online share recommendation and compare HGSRec with XGBoost via online A/B testing. Online service need to satisfy the following requirements: (我们在淘宝应用上部署HGSRec进行在线分享推荐，并通过在线A/B测试将HGSRec与XGBoost进行比较。在线服务需要满足以下要求：)
  • (1) Storage and processing for massive data. Share recommendation system is stored on MaxCompute as adjacency list for memory efficiency. (海量数据的存储和处理。共享推荐系统作为邻接列表存储在MaxCompute上，以提高存储效率。)
  • (2) Abnormal share action. We filter abnormal share actions (e.g., a user shares more than thousands of items with his friend within 24 hours). (异常的分享行为。我们过滤异常共享行为（例如，一个用户在24小时内与他的朋友共享超过数千个项目）。
  • (3) New feature and missing feature. New features comes everyday, so we leverage hash function to map all features, leading a slight loss of performance when hash collision happens. Missing features are padded with a specific $token$. （新功能和缺少的功能。新特性每天都会出现，所以我们利用哈希函数来映射所有特性，当哈希冲突发生时，会导致性能的轻微损失。缺少的功能用特定的标记填充。）
• (2) The online results range from 2020/01/08 to 2020/02/02 (25 days) are shown in Figure 7. Here we select UCTR (UCTR=Unique Click/Unique Visitor) for online evaluation. The larger UCTR, the better performance. The long-term observations show that HGSRec consistently outperforms XGBoost with a significant gap, demonstrating the high industrial practicability and stability of HGSRec. (在线结果范围从2020/01/08到2020/02/02（25天），如图7所示。在这里，我们选择UCTR（UCTR=Unique Click/Unique Visitor）进行在线评估。UCTR越大，性能越好。长期观察表明，HGSRec始终优于XGBoost，存在显著差距，表明HGSRec具有高度的工业实用性和稳定性。)
  

6 Conclusion

• (1) In this paper, we first study the problem of share recommendation in social e-commerce, whose goal is to predict whether a user will share an item to his friend. (在本文中，我们首先研究了社交电子商务中的共享推荐问题，其目标是预测用户是否会将一个项目共享给他的朋友。)
• (2) We first construct an attributed heterogeneous graph to represent share scenario (我们首先构造一个属性异构图来表示共享场景)
• (3) and propose a novel heterogeneous GNN based share recommendation model, called HGSRec. (提出了一种新的基于异构GNN的共享推荐模型HGSRec)
  • With the help of feature embedding and semantic aggregation, the proposed HGSRec learns multiple embeddings of $u,i,v$ under different meta-paths via tripartite heterogeneous GNNs, (在特征嵌入和语义聚合的帮助下，HGSRec通过三个异构GNN学习不同元路径下$u,i,v$的多个嵌入)
  • and then dynamically fuses them via dual co-attention mechanism, (然后通过双重共同注意机制将它们动态融合)
  • followed by a transitive triplet representation to model the asymmetric share action. (然后是一个传递的三重态表示来模拟不对称的共享行为。)
• (4) Extensive experiments on Taobao demonstrate the superiority of the proposed HGSRec.

Acknowledgements

References