Created: Jul 4, 2020 3:07 PM
Tags: KDD`19, 注意力机制, 知识图谱
1 Target
most existing works are unaware of the relationships between these entities and users.
we investigate how to explore these relationships which are essentially determined by the interactions among entities.
2 Model
将实体间交互分为两种类型：实体间交互和实体内交互。inter-entity-interaction and intra-entity-interaction.
(AKUPM) ,propose a novel model named Attention-enhanced Knowledge-aware User Preference Model ****
a self-attention network to capture the inter-entity-interaction by learning appropriate importance of
each entity w.r.t the user.
the intra-entity-interaction is modeled by projecting each entity into its connected relation spaces to obtain the suitable characteristics
- click-through rate (CTR) prediction 点击率预测
3 Contributions
- We propose a knowledge-aware model to alleviate the sparsity problem in recommendation systems based on a novel adaptive self-attention modeling design.
- this is the first work fully exploring the relationships between users and incorporated entities based on categorizing the interactions of entities into two types: intra-entity-interaction and inter-entity-interaction.
- AKUPM is able to figure out the most related part of the incorporated entities for each user, so as to make better CTR predictions. 2 real-world public datasets
INTRODUCTION
基于协同过滤的方法有很长的历史，但在矩阵稀疏时不好用，需要添加辅助的补充信息，近年来，知识图谱脱颖而出。
knowledge graph
A knowledge graph is a type of multi-relational directed graph composed of a large number of entities and relations. More specifically, each edge in the knowledge graph is represented as a triple in the form of (head entity, relation, tail entity), also called a fact, indicating that the head entity and the tail entity is connected through the relation
最左是用户，最右是实体entity
- Inter-entity-interaction: the importance of an entity varies a lot when included in different entity sets due to the interactions among entities. As shown in Figure 1, with regard to Bob, all incorporated movies are from USA, whereas to Steph the incorporated movies are from various countries. In this case, USA is of great importance to identify Bos’s interests rather than to identify Steph’s interests.由于实体之间的相互作用，当实体包含在不同的实体集合中时，其重要性会有很大的变化。Bob喜欢的电影都是来自美国，而斯蒂芬喜欢的电影哪个国家都有，因此在判断Bob的喜好时，美国很重要，但对于斯蒂芬却不是很重要。
- Intra-entity-interaction: for a certain user, an entity may show different characteristics when involved in different relations. For example in Figure 1, Alice may like City Lights since Charles Chaplin is the director of this film, whereas she may like Modern Time since Charles Chaplin acts the leading role in this film.一个实体在不同的关系里展示不同的特性，Alice喜欢City Lights因为CC是导演，然而她也可能喜欢MT因为CC是主演
🍊CTR预测与推荐系统
CTR预估模型可以用于推荐系统，因为推荐系统把CTR预估模型产生的CTR值当作排序的依据
/
ctr预估只是推荐系统中的一环。
一般推荐系统包括召回，精排（ctr预估），rerank（机制策略）。召回和精排的打分集合是不一样的。召回针对的是全部item，而精排针对的是召回输出的item。因此召回一般是在全部item集合上构建训练样本，而精排一般是基于展现样本来构建训练样本，而这部分样本本身是有偏的。即使精排有能力对全体item进行打分，由于只基于展现样本训练，对于没有展现过的item，预估会有偏差，可能会有问题。因此现阶段需要依赖召回通过各种召回方式来对item进行过滤筛选。但是后面算力足够了，这个问题是不是就不能解了呢？我觉得未必不能解。只不过在算力还没达到的时候，大家的精力暂时不在这个地方。
另外推荐系统的排序目标一般是多种多样的，以阿里电商推荐为例，一个主要排序目标是gmv（ctr×cvr×price），同时要兼顾用户体验，考虑多样性等指标，这些仅仅靠一个ctr预估模型是无法做到的。
/
常见的工业级推荐系统一般包含两个主要部分：召回和 ranking。为什么会有召回呢，是因为在工业界，我们需要推荐的内容常常百万千万甚至亿级的规模，不经过召回初筛，直接取特征进行 ranking 预测耗时是不现实的。召回的种类大概可以分为几类：一大类是基于用户兴趣的召回，包括长期兴趣、实时兴趣等等，第二类是协同类召回，比如基于用户session 链的协同、基于用户社交关系应用于内容的协同等等，还有一类是 nn 学习的 embedding 相似召回。Ranking 是将各路召回数据整合之后基于一个或者多个特定目标的模型排序部分，最常见的就是以点击率作为目标来进行预测，而还有其他一些目标比如说：停留时长、点赞评论等等。这里边常用的模型包括 LR、GBDT、DNN及各种改进的深度学习模型等等，当然还有多目标的模型，比如阿里的 ESMM。
