FM-Intent: Predicting User Session Intent with Hierarchical Multi-Task Learning
FM-Intent is a novel hierarchical multi-task learning model developed by Netflix to enhance recommendation systems by predicting user intent, significantly improving next-item prediction accuracy and offering personalized user experiences.
MAIN POINTS
- FM-Intent improves recommendation accuracy by predicting user intent using hierarchical multi-task learning.
- It leverages both short-term and long-term implicit signals to capture user session intent.
- The model demonstrates a 7.4% improvement over state-of-the-art models in next-item prediction accuracy.
- FM-Intent's intent predictions enhance personalized UI, analytics, and search optimization.
TAKEAWAYS
- FM-Intent integrates user intent prediction into Netflix's recommendation system, enriching user experience.
- The model employs a Transformer encoder for effective long-term interest modeling.
- Intent embeddings enable user clustering, revealing distinct viewing patterns.
- FM-Intent's hierarchical approach informs next-item recommendations, enhancing model coherence and effectiveness.