A Personalized Music Recommendation Algorithm Based on User Implicit Feedback and Weighted Socialized Tag Content Filtering

Yang Yang, Qiang WangProceedings of the 4th International Conference on Education, Management, Arts, Economics and Social Science (ICEMAESS 2017) 2017 · DOI

Abstract

The popularity of the Internet and the electronization of music resources make it easier for people to get music what they like. However, facing abundant resources, it is difficult for people to find their favorite music accurately and timely. Personalized music recommendation algorithms play an increasingly important role in online music service systems. Traditional feedback method requires the user to give feedback explicitly, which not only increases information collection cost but also gives the user an additional burden. This paper proposes a content filtering recommendation method based on user implicit feedback and weighted socialized tags and designs experiment to compare our recommendation method with fuzzy theory method. The experimental results show this method is more accurate than the fuzzy theory method in recommending music to the user.

Introduction

Modern music recommendation systems often struggle with the "user burden" associated with explicit feedback mechanisms, such as star ratings or manual tagging. To address this, researchers are increasingly looking toward implicit feedback—data derived from natural user interactions like skipping, repeating, or full-track listening. However, implicit feedback is inherently noisy; a skip might indicate a dislike or simply an external interruption.

This paper introduces a novel music recommendation algorithm that leverages implicit feedback mapped to "socialized tags" (user-generated labels from platforms like Last.fm). By creating a weighted content-filtering model, the authors translate raw behavioral logs into a high-fidelity user interest model, significantly improving recommendation precision over traditional fuzzy-logic approaches.

Methodology

The proposed system architecture consists of three primary modules: User Behavior Analysis, User Interest Modeling, and Filtering Recommendation.

System architecture of the proposed recommendation algorithm.

1. User Behavior Analysis

The system categorizes user interaction logs into three distinct sets based on interaction duration and frequency:

  • Skip Set: Tracks listened to for less than 50% of their duration, indicating potential disinterest.
  • Repeat Set: Tracks played multiple times or on loop, indicating strong preference.
  • Normal Set: Tracks played once in their entirety without further interaction.

2. User Interest Modeling

Rather than relying on complex acoustic features (e.g., pitch, tempo), the system utilizes "socialized tags"—crowdsourced semantic labels. The system extracts the most frequent tags for songs in each behavioral set to create three tag vectors: Positive Tags (PT), Normal Tags (NMT), and Negative Tags (UT).

3. Weighted Filtering

A preference score ($r_{ui}$) is calculated by linearly combining these vectors using three parameters ($\alpha, \beta, \gamma$). This weighting allows the algorithm to amplify tags associated with preferred music while penalizing those associated with skipped tracks.

Experimental Setup

The study utilized the Last.fm Dataset-1Kusers, containing approximately 19.2 million records from 992 users. To supplement the raw data, a web crawler was employed to retrieve track durations (to identify skips) and social tags via the track.info API. The data was split chronologically, with the first two-thirds used for training and the remaining third for testing.

Results and Discussion

The proposed model was evaluated against a baseline "Fuzzy Theory" content-filtering method. The primary metric for comparison was precision across varying lengths of the recommendation list ($k$).

Precision comparison between the proposed model and the fuzzy theory baseline across various recommendation list lengths (k).

Key findings include:

  • Superior Precision: The proposed model consistently outperformed the fuzzy theory method. At $k=25$, the model achieved a precision of 0.976, compared to 0.96 for the baseline.
  • Average Improvement: Across all tested values of $k$, the model showed an average precision increase of 1.68%.
  • Semantic Advantage: The results suggest that social tags effectively capture the "context" and "feel" of music in a way that aligns more closely with human perception than raw acoustic data.

Conclusion and Future Work

The research demonstrates that weighted socialized tags combined with implicit feedback provide a robust framework for personalized music recommendation. By interpreting natural listening behaviors, the system reduces user burden while maintaining high accuracy.

Future work will focus on automating the optimization of weighting parameters ($\alpha, \beta, \gamma$) for individual users, addressing the "cold start" problem for untagged new releases, and incorporating session-based context to account for shifting user moods.

Summary updated

Abstract

The popularity of the Internet and the electronization of music resources make it easier for people to get music what they like. However, facing abundant resources, it is difficult for people to find their favorite music accurately and timely. Personalized music recommendation algorithms play an increasingly important role in online music service systems. Traditional feedback method requires the user to give feedback explicitly, which not only increases information collection cost but also gives the user an additional burden. This paper proposes a content filtering recommendation method based on user implicit feedback and weighted socialized tagsand designs experiment to compare our recommendation method with fuzzy theory method. The experimental results show this method is more accurate than the fuzzy theory method in recommending music to the user.
BibTeX
@inproceedings{yang2017personalized,
  author = {Yang Yang and Qiang Wang},
  title = {{A Personalized Music Recommendation Algorithm Based on User Implicit Feedback and Weighted Socialized Tag Content Filtering}},
  booktitle = {Proceedings of the 4th International Conference on Education, Management, Arts, Economics and Social Science (ICEMAESS 2017)},
  year = {2017},
  doi = {10.2991/icemaess-17.2017.69},
}
Figure