Recsys Challenge 2018

比赛链接

赛题原文 This year’s challenge focuses on music recommendation, specifically the challenge of automatic playlist continuation. By suggesting appropriate songs to add to a playlist, a Recommender System can increase user engagement by making playlist creation easier, as well as extending listening beyond the end of existing playlists.

As part of this challenge, Spotify will be releasing a public dataset of playlists, consisting of a large number of playlist titles and associated track listings. The evaluation set will contain a set of playlists from which a number of tracks have been withheld. The task will be to predict the missing tracks in those playlists.

赛题理解

题目分析

Spotify提供了1M个播放列表,每个播放列表有部分隐藏,为每个给定的播放列表推荐500首歌。同时,这些播放列表各对应了1个子任务,共有10个子任务。由于所给数据集并没有提供用户的偏好,或者可以提供用户个性的数据,该任务可以看做是基于Session的推荐系统。每个播放列表就是一个特定的session,推荐系统需要根据这个session完成next-track推荐。

评价指标

比赛提供了在3个评价指标

  • Precision@N: N表示某个播放列表中隐藏的歌曲数,该评价指标计算所提供的500个歌曲命中N的比例。
  • NDCG@500: 该评价指标考虑了推荐列表中各item的位置,不难理解如果提供的500个歌曲的排序完美匹配,那么ndcg的取最大值1
  • clicks@500: 这个是一个基于真实使用场景的有趣的评级指标,可以想象现在用户正在查看推荐的500条歌曲,由于前端采用分页的形式,每页10条,共50页,如果用户发现当前页没有喜欢的歌曲就翻页,click自增1,当用户发现这500条没有喜欢的歌曲时,点击次数为51次,因此clicks@500的最大值是51。如果用户在某页匹配到感兴趣的歌曲,那么停止计数。

方法分类

  • 基于kNN的方法: 此类方法效率高,不利来GPU等高性能硬件,可以在短时间内完成训练,且准确率方面也较高,适合于当前工业界实践。
  • 基于神经网络的方法: 基于神经网络的方法优势在于拟合能力强,在数据量增大的情况下,相对于传统方法拥有更强的性能。同时,对于音乐的拓扑结构、音乐图片、标题、音乐内容等元数据的信息的提取具有显著优势。但缺点在于参数搜索空间大,不利于上线测试,需要的计算机资源也是另一个问题。
  • 其他方法

论文列表

感谢github项目RecSys-for-Playlist-Continuation对该比赛workshop论文及最新论文的整理
下面对各论文进行摘录,并补充个人理解的链接。

2018

1) Neighborhood-based approaches:

  • [Paper] [Code] [论文阅读]Efficient K-NN for Playlist Continuation (RecSys'18 Challenge)
  • [Paper] [Code] Effective Nearest-Neighbor Music Recommendations (RecSys'18 Challenge)
  • [Paper] [Code] [论文阅读]Automatic Music Playlist Continuation via Neighbor-based Collaborative Filtering and Discriminative ReweightingReranking (RecSys'18 Challenge)
  • [Paper] [Code] Efficient Similarity Based Methods For The Playlist Continuation Task (RecSys'18 Challenge)

2) Different approaches:

  • [Paper] [Code] Automatic Playlist Continuation using Subprofile-Aware Diversification (RecSys'18 Challenge)
  • [Paper] [Code] Random Walk with Restart for Automatic Playlist Continuation and Query-Specific Adaptations (RecSys'18 Challenge)
  • [Paper] [Code] Automatic playlist continuation using a hybrid recommender system combining features from text and audio (RecSys'18 Challenge)
  • [Paper] [Code] A Line in the Sand: Recommendation or Ad-hoc Retrieval? (RecSys'18 Challenge)

3) Neural network approaches:

  • [Paper] [Code] An Ensemble Approach of Recurrent Neural Networks using Pre-Trained Embeddings for Playlist Completion (RecSys'18 Challenge)
  • [Paper] [Code] Towards Seed-Free Music Playlist Generation Enhancing Collaborative Filtering with Playlist Title Information (RecSys'18 Challenge)
  • [Paper] [Code] TrailMix: An Ensemble Recommender System for Playlist Curation and Continuation (RecSys'18 Challenge)
  • [Paper] [Code] Using Adversarial Autoencoders for Multi-Modal Automatic Playlist Continuation (RecSys'18 Challenge)

4) Top-performing approaches:

  • [Paper] [Code] A hybrid two-stage recommender system for automatic playlist continuation (RecSys'18 Challenge)
  • [Paper] [Code] Artist-driven layering and user's behaviour impact on recommendations in a playlist continuation scenario (RecSys'18 Challenge)
  • [Paper] [Code] MMCF: Multimodal Collaborative Filtering for Automatic Playlist Continuation (RecSys'18 Challenge)
  • [Paper] [Code] Two-stage Model for Automatic Playlist Continuation at Scale (RecSys'18 Challenge)

2019

Conferences

  • [Paper] Offline Evaluation to Make Decisions About Playlist Recommendation Algorithms (WSDM'19)
  • [Paper] [Code] Adversarial Mahalanobis Distance-based Attentive Song Recommender for Automatic Playlist Continuation (SIGIR'19)
  • [Paper] [Code] Social Tags and Emotions as main Features for the Next Song To Play in Automatic Playlist Continuation (UMAP'19)

Journals

  • [Paper] A Hybrid Recommender System for Improving Automatic Playlist Continuation (TKDE'19)
  • [Paper] An Analysis of Approaches Taken in the ACM RecSys Challenge 2018 for Automatic Music Playlist Continuation (TIST'19)

2020

Conferences

  • [Paper] [Code] Consistency-Aware Recommendation for User-Generated ItemList Continuation (WSDM'20)
  • [Paper] User Recommendation in Content Curation Platforms (WSDM'20)
持续更新中...