Improving Community Detection Performance in Heterogeneous Music Network by Learning Edge-type Usefulness Distribution

Zheng Gao, Chun Guo, Xiaozhong Liu

Abstract

With music becoming an essential part of daily life, there is an urgent need to develop recommendation systems to assist people targeting better songs with fewer efforts. As the interactions between users and songs naturally construct a complex network, community detection approaches can be applied to reveal users' potential interests on songs by grouping relevant users & songs to the same community. However, as the types of interaction could be heterogeneous, it challenges conventional community detection methods designed originally for homogeneous networks. Although there are existing works on heterogeneous community detection, they are mostly task-driven approaches and not feasible for specific music recommendation. In this paper, we propose a genetic based approach to learn an edge-type usefulness distribution (ETUD) for all edge-types in heterogeneous music networks. ETUD can be regarded as a linear function to project all edges to the same latent space and make them comparable. Therefore a heterogeneous network can be converted to a homogeneous one where those conventional methods are eligible to use. We validate the proposed model on a heterogeneous music network constructed from an online music streaming service. Results show that for conventional methods, ETUD can help to detect communities significantly improving music recommendation accuracy while simultaneously reducing user searching cost.

Introduction

Modern music streaming services (MSS) generate massive amounts of data that can be modeled as heterogeneous networks. Unlike homogeneous networks where all nodes and edges represent the same entities and relationships, music networks involve diverse nodes (users, songs, artists, albums, playlists) and multiple interaction types (plays, bookmarks, comments, inclusions).

Traditional community detection algorithms, such as Louvain or Infomap, are typically designed for homogeneous graphs. When applied to heterogeneous music data, these methods often struggle because they treat all interactions as equally significant. For instance, a casual comment on a song might be weighted the same as a user playing a song hundreds of times. This paper addresses this gap by introducing a method to learn the relative "usefulness" of different edge types, effectively flattening a heterogeneous network into a weighted homogeneous one.

The ETUD Framework

The core of the proposed approach is the Edge-type Usefulness Distribution (ETUD). ETUD assigns a weight to each type of interaction in the network, reflecting its predictive power for user preferences. By learning these weights, the system can distinguish between high-intent behaviors (like repeated plays) and lower-intent interactions (like casual comments).

Community structure comparison showing how edge-type weights refine community boundaries.

To optimize these weights, the authors employ a Genetic Algorithm (GA). The GA treats a vector of weights as a "chromosome" and evolves it through several stages: initialization, evaluation (using Personalized PageRank and NDCG@10), selection, cross-over, and mutation. This evolutionary process continues until the weights stabilize, providing an optimal distribution that prioritizes high-signal behaviors.

Experimental Evaluation

The researchers evaluated ETUD using a dataset from Xiami Music, containing over 5.2 million edges across 16 types. They compared six baseline algorithms—including embedding-based (DeepWalk, LINE, node2vec), modularity-based (Louvain), and random walk-based (Infomap) methods—with and without the learned ETUD weights.

Performance comparison showing reduced search costs across all algorithms when ETUD weights are applied.

The results demonstrate a consistent improvement across two primary metrics:

  • Searching Cost: This metric measures the effort required for a user to find relevant songs within the detected communities. As shown in the performance table, applying ETUD weights significantly reduced the search cost for every algorithm tested. For example, the cost for the Infomap algorithm at the top 100 songs dropped from ~19,784 to ~15,997.
  • Recommendation Accuracy: Accuracy, measured by NDCG, also saw substantial gains. For instance, DeepWalk's accuracy rose from 0.6932 to 0.7529 when ETUD was integrated, and LINE's NDCG@5 jumped from 0.4883 to 0.5153.

Conclusion

By learning an Edge-type Usefulness Distribution, this research provides a robust bridge between complex heterogeneous data and established community detection algorithms. The genetic algorithm-based approach successfully identifies which user behaviors are most indicative of interest, leading to more cohesive communities and more accurate music recommendations. Future work may explore more dynamic weighting schemes or the use of Reinforcement Learning to further refine the ETUD estimation process.

Summary updated

Abstract

With music becoming an essential part of daily life, there is an urgent need to develop recommendation systems to assist people targeting better songs with fewer efforts. As the interactions between users and songs naturally construct a complex network, community detection approaches can be applied to reveal users' potential interests on songs by grouping relevant users & songs to the same community. However, as the types of interaction could be heterogeneous, it challenges conventional community detection methods designed originally for homogeneous networks. Although there are existing works on heterogeneous community detection, they are mostly task-driven approaches and not feasible for specific music recommendation. In this paper, we propose a genetic based approach to learn an edge-type usefulness distribution (ETUD) for all edge-types in heterogeneous music networks. ETUD can be regarded as a linear function to project all edges to the same latent space and make them comparable. Therefore a heterogeneous network can be converted to a homogeneous one where those conventional methods are eligible to use. We validate the proposed model on a heterogeneous music network constructed from an online music streaming service. Results show that for conventional methods, ETUD can help to detect communities significantly improving music recommendation accuracy while simultaneously reducing user searching cost.
BibTeX
@inproceedings{gao2022improving,
  author = {Zheng Gao and Chun Guo and Xiaozhong Liu},
  title = {{Improving Community Detection Performance in Heterogeneous Music Network by Learning Edge-type Usefulness Distribution}},
  booktitle = {Iconference},
  year = {2022},
  doi = {10.1007/978-3-030-96960-8\_5},
}
Figure