Automatic Feature Generation on Heterogeneous Graph for Music Recommendation

Chun Guo, Xiaozhong LiuSIGIR 2015 · DOI

TL;DR

This paper tackles the challenge of music recommendation by moving beyond traditional collaborative filtering and content-based methods. They propose a novel approach using heterogeneous graph mining to automatically generate rich "meta-path" features from diverse user and music data, significantly improving recommendation accuracy.

What Problem Does This Paper Address?

Online music streaming services (MSS) have exploded, amassing vast music libraries and, crucially, enormous amounts of user-generated data. This data, ranging from listening history and comments to user-created playlists, offers a rich tapestry of information that could enhance music recommendations. However, most existing recommendation systems either rely solely on audio content features (Content-Based Model - CBM) or simple user-music interactions like ratings and listening history (Collaborative Filtering - CF). This overlooks the complex, multi-faceted relationships present in user data. The authors identify a gap in leveraging this heterogeneous data effectively, proposing a method to model and exploit these intricate connections for better music recommendations.

Methods

The core of this paper lies in heterogeneous graph mining. They construct a rich graph where nodes represent different entities (users, artists, albums, songs, playlists, genres) and edges represent various relationships (e.g., "user plays song," "album includes song," "artist categorized as genre").

The key innovation is the automatic generation of meta-path based features. A meta-path is a sequence of node and edge types that describes a specific type of relationship between two nodes. For instance, a meta-path like "User -> Playlist -> Song" captures the relationship between a user and a song through their shared playlist. Instead of manually defining these paths (which can be overwhelming given the graph's complexity), the authors use a graph schema mining approach to identify all possible meta-paths up to a certain length.

These meta-path features are then used in a learning-to-rank (LTR) framework. For a given user, they calculate ranking scores for candidate songs based on these meta-path features. An LTR model is trained to combine these features and produce a final ranking of recommended songs. This approach allows them to learn which meta-paths are most predictive of user preference.

Datasets

The authors used data from Xiami.com, a large online music streaming service in China. The dataset includes:

  • 56,055 artists
  • 43,086 albums
  • 1,233,651 songs
  • 633 genres
  • 677,275 users
  • 305,916 playlists

This data was used to construct a heterogeneous graph with 15,929,369 edges representing various relationships. For evaluation, they selected 1,000 users with at least 50 songs in their listening history and randomly removed 40% of their "play" relationships to simulate a recommendation scenario.

Key Insights and Ideas

  • Heterogeneity is Key: The paper strongly advocates for leveraging the diverse types of relationships present in user-generated data, rather than relying on a single type of interaction.
  • Meta-Paths as Rich Features: Automatically generated meta-paths provide a powerful way to capture complex, indirect relationships that simple CF or CBM methods miss.
  • Schema Mining for Feature Generation: Exhaustively exploring the graph schema to identify potential meta-paths is crucial for discovering effective recommendation signals.
  • Learning to Rank for Integration: An LTR approach is essential for effectively combining the signals from multiple meta-path features and optimizing the final recommendation ranking.
  • Artist, Album, and User Relationships Matter: The best-performing meta-paths often involve user, artist, and album nodes, suggesting these entities and their connections are highly indicative of user preferences.

Key Results

The proposed heterogeneous graph mining approach, combined with learning to rank, significantly outperformed a baseline item-to-item collaborative filtering algorithm.

  • The LTR model using meta-path features achieved substantial improvements in metrics like MAP@10, MAP@100, NDCG@10, and NDCG@100, with statistical significance (p < 0.0001).
  • Specifically, the LTR model with meta-paths of length < 3 combined with CF achieved a MAP@10 of 0.0069, a significant jump from the CF baseline's 0.0044.
  • While single meta-path ranking functions showed promise, especially for ranks beyond 100, the LTR approach integrating multiple meta-paths proved superior.
  • The best performing meta-path identified was UP Ar Pin Al S?, suggesting a path from User to Playlist, then to Artist, then to Album, and finally to Song.

Limitations and Future Work

  • Meta-Path Selection: While they automatically generate meta-paths, the paper notes that the "best performing meta-paths are more likely to be related to artist (Ar), album (Al) and user (U)." This suggests that not all meta-paths are equally informative, and further research into more sophisticated feature selection methods is warranted.
  • Other Recommendation Tasks: The authors propose extending their work to other recommendation tasks, such as recommending music to playlists or recommending existing playlists to users.
  • Sophisticated Feature Selection: Investigating more advanced techniques for selecting the most impactful meta-path features is a key future direction.

Summary updated

BibTeX
@inproceedings{guo2015automatic,
  author = {Chun Guo and Xiaozhong Liu},
  title = {{Automatic Feature Generation on Heterogeneous Graph for Music Recommendation}},
  booktitle = {International ACM SIGIR Conference on Research and Development in Information Retrieval},
  year = {2015},
  doi = {10.1145/2766462.2767808},
}
Figure