Enhancing Sequential Music Recommendation with Personalized Popularity Awareness

Abstract: In the realm of music recommendation, sequential recommender systems have shown promise in capturing the dynamic nature of music consumption. Nevertheless, traditional Transformer-based models, such as SASRec and BERT4Rec, while effective, encounter challenges due to the unique characteristics of music listening habits. In fact, existing models struggle to create a coherent listening experience due to rapidly evolving preferences. Moreover, music consumption is characterized by a prevalence of repeated listening, i.e., users frequently return to their favourite tracks, an important signal that could be framed as individual or personalized popularity. This paper addresses these challenges by introducing a novel approach that incorporates personalized popularity information into sequential recommendation. By combining user-item popularity scores with model-generated scores, our method effectively balances the exploration of new music with the satisfaction of user preferences. Experimental results demonstrate that a Personalized Most Popular recommender, a method solely based on user-specific popularity, outperforms existing state-of-the-art models. Furthermore, augmenting Transformer-based models with personalized popularity awareness yields superior performance, showing improvements ranging from 25.2% to 69.8%. The code for this paper is available at https://github.com/sisinflab/personalized-popularity-awareness.

• The paper presents a novel method that integrates personalized popularity signals into sequential music recommendation models.
• It enhances Transformer-based approaches by adjusting logit computations, resulting in performance improvements ranging from 25.2% to 69.8%.
• Experimental results on Yandex Music and Last.fm datasets confirm that accounting for user-specific replay patterns significantly optimizes recommendation relevance.

Enhancing Sequential Music Recommendation with Personalized Popularity Awareness

This paper, authored by Davide Abbattista et al., addresses noteworthy challenges in the domain of music recommender systems by proposing a novel methodology that integrates personalized popularity signals into sequential recommendation models. The work specifically targets the intrinsic characteristics of music consumption, such as dynamic user preferences and repeated listening behaviors, which traditional Transformer-based models like SASRec and BERT4Rec struggle to capture effectively.

Key Contributions

1. Novelty in Approach: The paper introduces an innovative method to incorporate personalized popularity information directly into sequential recommendation models. This approach balances the need for exploring new music and gratifying user preferences for familiar tracks.
2. Experimental Evidences: Extensive experiments demonstrate that the Personalized Most Popular recommender outperforms existing state-of-the-art models. Additionally, Transformer-based models augmented with personalized popularity scores show significant performance improvements.

Methodology

The core premise of this study lies in leveraging personalized popularity—i.e., user-specific repeated listening patterns—as a signal to enhance sequential music recommendations. The methodology entails modifying the probability distribution calculations within Transformer-based models to integrate these personalized popularity scores.

Implementation for Softmax-based Models

For BERT4Rec, which utilizes the softmax function, personalized popularity probabilities are transformed into logit scores that can be combined with the model's logits.

• Personalized Popularity Probability:

$$\hat{p}_P(j)=\frac{\frac{c_j+\epsilon}{\max(C+\epsilon)}}{\sum_{z=1}^{N} \frac{c_z+\epsilon}{\max(C+\epsilon)}}$$

• Personalized Popularity Score:

$$y_j = \ln \left( \frac{c_j + \epsilon}{\max(C + \epsilon)} \right)$$

These scores are then summed with the initial scores from BERT4Rec to adjust item rankings.

Implementation for Sigmoid-based Models

For models that utilize the sigmoid function, such as SASRec and gSASRec, the personalized popularity scores are obtained by transforming the probabilities into a form suitable for sigmoid activation.

• Personalized Popularity Score:

$$y_j = - \ln \left( \frac{1 - \hat{p}_P(j)}{\hat{p}_P(j)} \right)$$

Experimental Setup and Results

The experiments leverage two music datasets: Yandex Music Event and Last.fm-1K. The models evaluated include Most Popular, Personalized Most Popular, BERT4Rec, SASRec, and gSASRec, both with and without the integration of personalized popularity scores.

Key Findings:

1. Baseline Superiority: The Personalized Most Popular recommender exhibited superior performance over state-of-the-art models such as BERT4Rec, SASRec, and gSASRec, particularly highlighting the significance of repeated consumption patterns in music recommendation.
2. Enhanced Performance with PPS: Integrating personalized popularity scores into Transformer-based models (BERT4Rec, SASRec, gSASRec) resulted in substantial performance gains. Improvements ranged from 25.2% to 69.8% across various metrics like NDCG@5, NDCG@10, NDCG@40, and NDCG@100.

Implications and Future Directions

The research underscores the importance of acknowledging and integrating user-specific repeated interaction patterns to improve recommendation relevance. This integration not only advances the accuracy of existing models but also aligns predictions more closely with user listening habits.

Practical Implications:

• Improved User Satisfaction: By effectively incorporating users' preferences for replaying favorite tracks, recommender systems can maintain user engagement and satisfaction.
• Exploration vs. Exploitation: The balance achieved between recommending new and familiar tracks ensures a richer and more satisfying user experience.

Theoretical Implications:

• Model Adaptation: This study highlights the necessity for sequential models to adapt their learning processes to account for domain-specific behaviors, such as repeated consumptions in music.

Speculative Future Developments:

Future work could explore alternative integration techniques beyond direct score summation and assess the impact of personalized popularity in domains beyond music. Additionally, examining the long-term implications of popularity-aware recommendations, such as the risk of filter bubbles, would provide a comprehensive understanding of the approach's potential impacts.

Conclusion

This paper makes significant strides in enhancing music recommendation by addressing the limitations of existing Transformer-based models and introducing a personalized popularity-aware approach. The findings demonstrate that the integration of user-specific popularity scores can effectively improve the performance and relevance of music recommendation systems.