How Meta uses generative AI to understand user perception

Meta – parent company of Facebook, Instagram, WhatsApp, Threads and more – runs one of the largest recommendation systems in the world.

In two recently published papers, their researchers have shown how generative models can be used to better understand and respond to consumer intent.

By looking at proposals as a generational problem, you can approach it in new ways that are richer in content and more effective than classical methods. This approach can have important applications for any application that requires the retrieval of documents, products, or other materials.

Dense vs. generational recapture

The usual method of creation recommendation systems is the measurement, storage, and retrieval of dense representations of documents. For example, to recommend items to users, an application needs to train a model that can compute shelters for both users and objects. Then it needs to create a large repository of embedded objects.

At decision time, the recommender system tries to understand the user's intent by finding one or more items that have a similar setup to the user. This approach requires more storage and counting capacity as the number of objects grows because every input has to be stored and every recommendation function has to compare the user nesting and the entire source.

Generative discovery is a newer approach that tries to understand user intent and make recommendations by predicting the next thing in a sequence instead of searching database. Storing a grasp of objects does not require generational recapture and its resolution and storage costs remain constant as the list of objects grows.

The most important thing in doing generative discovery is to calculate “semantic IDs” (SIDs) that contain contextual information about each object. Generation replacement systems like TIGER work in two stages. First, an encoder model is trained to generate a unique input value for each object based on its description and properties. These input values ​​become SIDs and are stored with the object.

In the second stage, a Model transformers is trained to predict the next SID in an input sequence. The list of input SIDs represents the user's past interactions with objects and the model's prediction is the SID of the object to recommend. Generative recapture reduces the need for storage and analysis over the engineering of individual items. It also increases the ability to capture deeper semantic relationships within the data and provides other benefits of generative models, such as changing the temperature to change the diversity of propositions.

Advanced generation recall

Despite its lower storage and decision costs, there are some limitations to generation recovery. For example, he tends to over-fit the items he has seen during training, which means he has trouble dealing with items added to the catalog after the model has been built. to train. In recommendation systems, this is often referred to as “the cold start problem,” which applies to users and objects that are new and have no interaction history.

To address these shortcomings, Meta has developed a so-called hybrid recommendation system LIGERwhich combines the computational and storage efficiency of generative retrieval with robust ingest quality and dense retrieval ranking capabilities.

During training, LIGER uses both similarity score and benchmark objectives to improve the model's recommendations. During inference, LIGER selects a number of candidates based on the generation mechanism and supplements them with a few cold starts, which are then ranked based on the establishment of the generated candidates.

The researchers note that “the combination of dense and generative retrieval methods has great potential to improve recommendation systems” and as the models evolve, “they will become increasingly practical for applications in the -real world, enabling more personalized and meaningful user experiences.”

In a separate paper, the researchers introduce a new generation recruitment method called multimodality A multimodal choice view (Mender), an approach that enables generative models to construct interesting choices from user interactions with various objects. Mender builds on the generative recruitment methods based on SIDs and adds a few components that add suggestions with user preferences.

Mender uses a large-scale language model (LLM) to convert user interactions into specific options. For example, if the user has recommended or complained about a particular item in a review, the model will summarize it as a choice about that product category.

The main recommender model is trained to fit both a sequence of user interactions and the user's preferences when predicting the next semantic ID in the input sequence. in. This enables the recommendation model to learn in a general context and perform and adapt to user preferences without being explicitly trained on them.

“Our contributions pave the way for a new class of generative discovery models that unlock the potential of using organic data to drive recommendations through text users' preferences,” the researchers write.

Impact on enterprise applications

The effectiveness of generative recruitment systems can have a significant impact on enterprise applications. These advances translate into immediate practical benefits, including lower infrastructure costs and faster decision making. The technology's ability to keep storage and decision costs constant regardless of catalog size makes it especially valuable for growing businesses.

The benefits extend across industries, from e-commerce to enterprise search. Genre adoption is still in its early stages and we can expect applications and frameworks to emerge as it matures.