超图音乐推荐模型笔记

1 Title

        Music recommendation via hypergraph embedding（Valerio La Gatta，Vincenzo Moscato，Mirko Pennone，Marco Postiglione，Giancarlo Sperlí）【IEEE Transactions on Neural Networks and Learning Systems 2022】

2 Conclusion

               This paper proposes hypergraph embeddings for music recommendation (HEMR), a novel framework for song recommendation based on hypergraph embedding. The hypergraph data model allows us to represent seamlessly all the possible and complex interactions between users and songs with the related characteristics; meanwhile, embedding techniques provide a powerful way to infer the user–song similarities by vector mapping. 

3 Good Sentences

        1、The availability of massive multimedia content has brought to the need for such platforms to automatically provide each user with products that appeal to their own personal taste, needs, and preferences, rather than requiring them to browse huge catalogs before finding the right song or movie of interest. As a result, recommender systems have become an essential component of any streaming service [1], determining their success and the related fidelity of users.（The necessarity of the research of Recommender systems）
        2、In each of these types of recommender systems, the key challenge is to maximize the value of all the available information, especially when a new user is approaching the platform. To this end, graphs are a natural data structure to represent the heterogeneous information of users, songs, artists, and their interactions, and their value has been proven for general recommendation problems.（Why try to import hypergraph into Recommenderation problem）
        3、However, the usage of a hypergraph model still represents a novelty in the field, and while the results are very encouraging, there is still room for improvement. For example, as with every model-based strategy, the current iteration of the model fails at instantly returning embeddings for new users or songs, meaning that we first need to calculate all of the embeddings to return recommendations for these new items（The future developments of this study）

---

本文方法如图所示，大概有如上三个流程：

1、超图数据建模：数据在基于超图的结构中建模和存储。

2、嵌入生成：随机游走和顶点嵌入是从超图数据模型生成的。

3、推荐生产：为每个用户生成一首 top-K 歌曲推荐。

这个多阶段框架非常灵活，可以通过更改嵌入生成模块轻松改进，不仅可以用于推荐，还可以用于不同的应用，正好是我需要的。

在音乐推荐系统中，超图的点集被由以下这些内容组成：歌曲S，用户U，艺术家A，版本R，标签T。而边集根据点与点之间的关系，包含以下内容：收听(用户与歌曲)、专辑（版本与歌曲）、唱片（艺术家和歌曲）、主题（标签和歌曲）。

因此超图定义为：，生成的超图大概是这样的：