Unlocking User Intent with Meta’s AI Research

Meta runs one of the world’s largest recommendation algorithms across its social media channels. Recent research publications show how they use generative AI to understand user intent better and offer richer, more efficient recommendations, a technique that could help various retrieval applications.

Dense vs Generative Retrieval

Traditional recommendation systems depend on storing and comparing item embeddings to user preferences, which requires extensive storage and computation. In contrast, generative retrieval takes a simpler approach, expecting the next most likely item based on a user’s interaction history, reducing the need for vast embedding databases.

Here’s how it works:

Generative retrieval uses a two-phase approach: first creating unique semantic IDs (SIDs) for items using an encoder, then training a transformer to predict the next SID based on user history. This method maintains constant costs regardless of database size and offers better semantic understanding plus adjustable recommendation diversity.

Advanced Generative Retrieval

While generative retrieval is efficient, it has difficulty with new items and the cold start problem. LIGER, Meta’s solution, combines generative and dense retrieval techniques, using both prediction and similarity scores during training. At runtime, it generates ideas while simultaneously considering new items and ranking them based on embedding similarity.

Meta-researchers see huge opportunities in merging dense and generative retrieval methods. Their new Mender system improves on this method by using an LLM to extract user preferences from interactions and then combining these findings into suggestions. This enables the system to better understand and react to user preferences without requiring intentional training, possibly transforming the way recommendation systems learn from organic user input.

Implications for Enterprise Applications

Generative retrieval systems offer considerable business benefits due to their cost-effective scaling and quick performance. With constant storage and computation costs regardless of size, they’re especially useful for growing enterprises in e-commerce and enterprise search, though the technology is still in its early stages.

Source- VentureBeat