TLDR: SPARC is a new retrieval model for recommender systems that addresses limitations of existing multi-interest models. It uses a learnable, dynamic interest space via Residual Quantized Variational Autoencoder (RQ-VAE) and a probabilistic module for “soft-search,” enabling proactive exploration of user interests. This leads to better discovery of novel and long-tail content, improved user engagement, and better performance for new users, as validated by online A/B tests and offline evaluations.
Recommender systems are essential in today’s digital world, helping us discover new products, movies, music, and more. A key challenge for these systems is accurately understanding and adapting to our diverse and evolving interests. Traditional methods often struggle with this, leading to recommendations that can feel repetitive or miss out on new and niche content.
A new research paper introduces a novel framework called SPARC, which stands for Soft Probabilistic Adaptive multi-interest Retrieval Model via Codebooks. This model aims to overcome three major limitations of existing multi-interest retrieval methods in recommender systems. First, many current systems rely on static interest representations that don’t change as user preferences evolve. Second, they often focus too much on existing interests, missing opportunities to explore new or less common (long-tail) interests. Third, these models can struggle to effectively identify interests for users with limited historical interactions, a problem known as the cold-start issue.
How SPARC Works
SPARC addresses these challenges through two main innovations. The first is its use of a Residual Quantized Variational Autoencoder (RQ-VAE) to create a dynamic, learnable space for user interests. Unlike previous approaches where interest definitions were fixed, SPARC integrates the RQ-VAE directly into the recommendation model. This allows the system to learn and adjust interest codes based on real-time user feedback, making them ‘behavior-aware’ and capable of evolving dynamically. This joint training ensures that the learned interests are highly relevant to the actual recommendation task.
The second key innovation is a probabilistic interest module. This module predicts a user’s probability distribution across the entire dynamic interest space. This enables a unique ‘soft-search’ strategy during online recommendations. Instead of passively matching users to items based on their most dominant interests, SPARC proactively explores a wider range of potential interests. This shifts the retrieval paradigm from ‘passive matching’ to ‘proactive exploration’, significantly enhancing the discovery of novel and diverse content.
Also Read:
- Enhancing Recommendations with Semantic Item Graphs and Noise Robustness
- DiMuST: A New Approach to POI Recommendation with Social and Spatial-Temporal Insights
Real-World Impact and Performance
The effectiveness and practical value of SPARC have been validated through extensive experiments. Online A/B tests conducted on an industrial platform with tens of millions of daily active users showed substantial improvements in key business metrics. For instance, there was a 0.9% increase in user view duration, a 0.4% increase in user page views (PV), and a remarkable 22.7% improvement in PV500, which measures how quickly new content reaches 500 page views within 24 hours. This last metric highlights SPARC’s ability to promote the discovery of new and long-tail content.
Offline evaluations on open-source Amazon Product datasets also demonstrated consistent improvements in metrics like Recall@K and Normalized Discounted Cumulative Gain@K (NDCG@K), which measure the relevance and ranking quality of recommendations.
Further analysis revealed that SPARC significantly boosts the retrieval of long-tail items, showing a 24.2% relative improvement in Recall@50 for these items compared to strong baselines. It also led to higher coverage and intra-list diversity, meaning users were exposed to a broader and more varied set of recommendations. The model’s design also proved particularly beneficial for ‘cold-start’ users—those with sparse interaction histories—showing an 11.84% improvement in NDCG@50 for this group, indicating better generalization capabilities.
In conclusion, SPARC represents a significant step forward in multi-interest retrieval for recommender systems. By creating dynamic, behavior-aware interest representations and employing a probabilistic exploration strategy, it moves beyond traditional limitations to offer more diverse, novel, and personalized recommendations, ultimately improving user experience and engagement. For more technical details, you can refer to the full research paper.
