The Challenge
JustWatch already had a functioning recommender system, but evolving user expectations, scale, and advances in machine learning made its limitations increasingly visible. The challenge was not starting from scratch, but upgrading a mission-critical system without sacrificing robustness, efficiency, or business impact.
- Cold-start limitations: The existing collaborative filtering approach relied heavily on historical user interactions. Newly added titles lacked sufficient data, making it difficult to recommend fresh content early—exactly when discoverability matters most.
- Popularity bias vs. taste relevance: Interaction-based models naturally favored popular titles. While effective at scale, this made it difficult to surface niche or “hidden gem” content that better matched individual user taste, limiting recommendation quality for power users.
- Efficiency at massive scale: Serving recommendations to millions of users from a very large title corpus pushed the existing online serving stack to its limits. With modern advances in vector search, there was clear potential to significantly reduce latency and infrastructure costs.
Our Approach
We partnered closely with JustWatch’s in-house engineers and data scientists to mature the recommender system while keeping it fully embedded in their own infrastructure and way of working.
- Build, Don’t Buy—With Acceleration: Recommendations are core to JustWatch’s value proposition. Rather than opting for an off-the-shelf solution, we supported a build-in-house strategy, accelerating development with expertise from multiple large-scale recommendation projects.
- Two-Tower Architecture for Taste-Based Recommendations: We redesigned the recommender around a flexible two-tower model. This allowed the system to learn not only from user–title interactions, but also from rich title metadata such as genres, themes, actors, release dates, plot descriptions, and even visual signals like posters.
- Mitigating Cold Start with Metadata: By leveraging metadata available at title release, new movies and series could be recommended immediately—long before interaction data accumulated. This improved early discoverability and accelerated feedback loops.
- Balancing Popular and Niche Content: Through systematic experimentation, we tuned the model to balance mainstream popularity with personalized niche recommendations, increasing the chance users discover content that truly matches their taste.
- Modern MLOps and Efficient Serving: We implemented modern MLOps practices, including model tracking, registration, automated retraining, and deployment. Recommendations were served using Approximate Nearest Neighbor search, significantly improving performance while reducing operational costs.
"Building inhouse, with the acceleration of a partner like Enjins that brought in expertise from mutliple recommendation projects, was our go-to strategy.
"
