| 摘要: |
| 近年来，人们对时尚穿搭有了更高的美学追求。因此，个性化互补服装推荐，即为用户推荐与他/她已购买服装相匹配的互补服装，逐渐吸引了学术界的广泛关注。个性化互补服装推荐不同于一般的推荐任务（如电影推荐），它推荐的服装需要满足两个条件：1）与目标服装搭配；2）满足用户偏好。因此，相关的现有方法主要是基于时尚单品的多模态数据，着力于建模单品与单品之间的兼容性交互和用户与单品之间的偏好交互，以实现个性化互补服装的推荐。这些方法的缺点主要在于它们将每一个单品-单品交互或者用户-单品交互看作一个独立的数据实例，而忽略了单品的属性知识以及时尚实体（即，用户、单品及属性）之间的高阶交互关系。事实上，与一个单品（如，上衣）搭配的所有互补单品（如，下衣）可能会共享某些相同的属性（如，颜色）；同时，具有类似品味的用户也可能倾向于选择具有类似属性特征的单品。显然，这些时尚实体之间的高阶关系蕴含了丰富的有关单品兼容性和用户偏好的协同信号，因而能够促进个性化互补服装推荐模型性能的提升。据此，本文构建了一个大规模协同时尚图谱，并基于图卷积神经网络（Graph Neural Networks，GNNs）来探索时尚实体之间的高阶关系，进而更好地实现个性化的互补服装推荐。具体地，本文提出了一个新颖的基于时尚图谱增强的个性化互补服装推荐模型（Fashion Graph-enhanced Personalized Complementary Clothing Recommendation），简称为FG-PCCR。FG-PCCR由两个关键的部分组成：独立的一阶交互建模和协同的高阶交互建模。一方面，独立的一阶交互建模模块基于视觉和文本模态数据，致力于通过神经网络和矩阵分解方法分别对单品-单品搭配交互和用户-单品偏好交互进行综合性建模。另一方面，协同的高阶交互模块基于构建的协同的时尚图谱，通过图神经网络利用信息传播机制来提取高阶的协同信号，进一步丰富用户和单品的向量表示。FG-PCCR模型能够有效整合时尚实体之间的复杂的高阶关系信息，用户和单品的表示学习，进而改进个性化互补推荐的效果。最后，对于给定的用户和目标上衣，我们能够得到推荐的下衣的个性化兼容性分数。另外，在真实的时尚数据集上做的大量实验，充分地验证了本文所提模型FG-PCCR相对于基准方法的优越性。 |
| 关键词: 个性化服装推荐 时尚兼容性建模 图神经网络 |
| DOI：10.19363/J.cnki.cn10-1380/tn.2021.09.14 |
| 投稿时间：2021-04-28修订日期：2021-08-09 |
| 基金项目:本课题得到CCF-百度松果基金资助。 |
|
| Fashion Graph-enhanced Personalized Complementary Clothing Recommendation |
| SHI Jinwan,SONG Xuemeng,LIU Zixin,NIE Liqiang |
| School of Computer Science and Technology, Shandong University, Qingdao 266237, China |
| Abstract: |
| Recent years, people have a higher aesthetic pursuit of fashion clothing matching. Therefore, personalized complementary clothing recommendation, that is, to recommend complementary clothes to users for matching their existing garments, has attracted increasing research attention. Different from the general recommendation tasks (e.g., movie recommendation), the recommended items in the context of personalized complementary clothing recommendation must meet two requirements:1) match with the given item and 2) satisfy with the user's preference. Accordingly, existing research efforts mainly focus on modeling the item-item compatibility interaction and the user-item preference interaction based on the multi-modal data of fashion items to recommend items. One deficiency is that they regard each item-item or user-item interaction as an independent data instance, while overlooking the attribute information of items and the high-order relations among fashion entities, i.e., users, items, and attributes. In fact, items (e.g., bottoms) that go well with one item (e.g., a top) are more likely to share certain underlying attribute patterns (e.g., color), while users with similar tastes tend to choose items with similar attributes. Obviously, these high-order relations among fashion entities convey much implicit collaborative signals towards the item compatibility modeling and user preference modeling, and are critical for personalized complementary clothing recommendation. In light of this, in this work, we build a large-scale collaborative fashion graph to investigate the utility of high-order relations among fashion entities based on the Graph Convolutional Neural Networks(GNNs), in the context of personalized complementary clothing recommendation. In particular, we propose a new fashion graph-enhanced personalized complementary clothing recommendation model, dubbed as FG-PCCR, which consists of two key components:the independent one-order interaction modeling and the collaborative high-order interaction modeling. On the one hand, the independent one-order interaction modeling module, based on visual and textual modal data, is dedicated to the comprehensive modeling of item-item matching interaction and user-item preference interaction respectively by using neural network and matrix decomposition methods. On the other hand, the collaborative high-order interaction module is based on the constructed collaborative fashion graph, and distills the higher-order collaborative signals through the Graph Neural Networks using the information transmission mechanism to further enrich the vector representation of users and items. The FG-PCCR can effectively integrate the complex high-order relational information between fashion entities, the representation learning of users and items, and then improve the modeling effect of personalized clothing compatibility. Finally, for a given user and target top, the personalized compatibility score of the matching undergarment is obtained. In addition, extensive experiments on the real-world dataset have demonstrated the superiority of the proposed scheme over the state-of-the-art methods. |
| Key words: personalized fashion recommendation fashion compatibility modeling graph neural networks |
