Meta has presented a new paradigm for AI-based recommendation algorithms by introducing a new recommendation system using generative AI, ‘LIGER’, and a multimodal recommendation technology, ‘Mender’. This is the first case of combining generative AI with existing recommendation algorithms, and focuses on more precisely understanding user intent and improving recommendation quality.
|
▲[Korean Today] Meta’s newly developed recommendation system ‘Liger’ © Reporter Byun A-rong |
Existing recommendation systems converted document information into numbers, stored them, and operated in a dense search method that compared user requests with a large list of items. However, as the number of items increased, the amount of calculation and storage space increased exponentially, which was a limitation.
To solve this problem, Meta applied the ‘generative retrieval’ technology based on generative AI. Generative retrieval is a method of predicting the next recommended item based on data that the user has interacted with in the past. It can identify the user’s intention without searching the data, which can greatly reduce storage space and computational costs.
Meta presented ‘Liger’, a hybrid recommendation system that combines the advantages of generative search with the strengths of conventional dense retrieval.
Liger trains models for both generative search and dense search simultaneously during the training phase, maximizing recommendation quality by using similarity scores and next token predictions. In the inference phase, it uses a generative mechanism to select candidate items, and then adds new items (cold start items) to generate the final recommendation list.
Through this, Liger succeeded in improving user experience by combining computational efficiency and high embedding quality. Meta said, “Liger will increase the efficiency and accuracy of recommendation systems while providing more personalized experiences.”
Meta also unveiled ‘Mender’, a multimodal technology that improves recommendation quality by identifying preferences revealed as users interact with various items.
When a user shows a positive or negative reaction to a specific item, Mender converts this into a preference through a large language model (LLM). The model learned in this way can reflect the user’s preference in real time and provide more precise recommendations.
Meta’s new technology suggests the direction for the development of recommendation systems and is expected to enable more personalized AI-based user experiences. The research team emphasized that “Liger and Mender are the first steps toward opening up new possibilities for generative AI,” and that “recommendation systems will become the central axis of AI technology.”
These technologies, which provide users with customized content through the development of recommendation algorithms, are expected to bring about major changes in various fields such as social media, e-commerce, and content streaming.
© The Korean Today – All Rights Reserved