AI Engineer World’s Fair 2025 - LLM RECSYS

Introduction to Recommendation Systems and LLMs 02:38

• The speaker introduces the inaugural Rexus track at the AI Engineer World's Fair, focusing on merging recommendation systems with language models.
• Historical context notes the initial use of language modeling techniques in recommendation systems starting around 2013, evolving through various technologies like recurrent neural networks and transformers.

Key Ideas in Recommendation Systems 03:54

• The speaker presents three concepts: semantic IDs, data augmentation, and unified models.
• Discusses the challenges of hash-based item IDs, co-star problems, and sparsity in recommendations, advocating for semantic IDs that incorporate multimodal content.

Case Study: Quao and Semantic IDs 04:47

• Introduces Quao, a short video platform, and their method for using trainable multimodal semantic IDs to improve recommendations.
• Highlights improvement metrics, including a 3.6% increase in co-star coverage and enhanced recommendation quality.

Data Augmentation with Language Models 08:09

• Emphasizes the importance of data quality for machine learning, particularly in search and recommendations.
• Discusses examples from Spotify and Indeed demonstrating how LLMs can improve data quality and recommendation relevance.

Case Study: Indeed’s Job Recommendation System 09:07

• Indeed faced challenges with ineffective job recommendations leading to user unsubscriptions.
• They developed a lightweight classifier to filter bad job recommendations, eventually increasing application rates by 4% and reducing unsubscribe rates by 5%.

Case Study: Spotify’s Query Recommendation System 13:15

• Discusses Spotify's approach to introducing podcasts and audiobooks, utilizing query recommendations to enhance user discovery.
• Reports a 9% increase in exploratory queries as users engage with new content.

Unified Models in Recommendation Systems 16:13

• Addresses the inefficiencies of having multiple separate models for recommendations and search systems.
• Introduces the concept of unified models that improve operational efficiency and learning transfer across different tasks.

Case Study: Netflix’s Unified Contextual Ranker 17:16

• Netflix developed a unified ranker to streamline their recommendation systems, aiming to reduce operational costs and improve performance across multiple tasks.

Case Study: Etsy’s Unified Embedding Approach 19:17

• Discusses Etsy’s method for improving search results through unified embeddings that consider user preferences, leading to increased conversion rates.

Conclusion: Takeaways and Future Directions 22:24

• The speaker summarizes the key takeaways as semantic IDs, data augmentation, and unified models for enhancing recommendation systems.
• Encourages attendees to explore the rest of the talks in the track for deeper insights.

Audience Engagement and Questions 23:22

• The speaker invites questions from the audience, discussing various challenges and considerations in recommendation systems and LLMs, including the complexity of balancing different model capabilities and the practicalities of implementing these systems in real-world scenarios.