Spotify has announced its expansion into the audiobook market, bringing its vast collection of music and talk shows to a wider audience. However, the move poses challenges, particularly in regards to providing personalized audiobook recommendations. Since users cannot preview audiobooks in the same way they can music tracks, creating accurate and relevant recommendations is crucial. The process is further complicated by the issue of sparse data when incorporating a new content form into an existing platform.
To overcome these complications, a team of researchers has developed a new recommendation system, called 2T-HGNN. This system incorporates users’ current musical and podcast interests and uses a Two Tower (2T) architecture, combined with a Heterogeneous Graph Neural Network (HGNN), to expose complex connections between objects with low latency and complexity.
A component of this system’s design separates users from the HGNN graph, which allows for a deeper analysis of item relationships. A multi-link neighbour sampler has been added to enhance the recommendation process. These calculated decisions, combined with the 2T component, greatly reduce the computational complexity of the HGNN model.
Experiments with millions of users have proven the effectiveness of this approach. The 2T-HGNN model resulted in a substantial 23% rise in streaming rates and a 46% increase in the initiation rate of new audiobooks.
The research team highlighted their key contributions. They conducted a comprehensive study to understand user patterns in consuming audiobooks. They also proposed a modular architecture, which combines the 2T model and HGNN into a single stack, making it easier to integrate audiobooks into the existing Spotify platform.
The researchers also addressed data distribution imbalance by integrating an innovative edge sampler into the HGNN. They argued that effective user-audiobook predictions can be generated by integrating weak signals in user representation.
The 2T-HGNN model has passed extensive offline trials, consistently outperforming other models. The findings of these trials were further validated by positive feedback from users who participated in A/B tests.
In conclusion, by employing user preferences, advanced graph-based strategies, and efficient computational methodologies, the team developed a novel recommendation system for integrating audiobooks into the Spotify platform. This move could significantly enhance user experience and contribute to the overall diversity of the digital audio environment.
