Here’s the translation into American English:
—
Recommendation systems face a challenge known as the “cold start,” which manifests not only in the onboarding of new users or items but also in the lack of personalized signals from the outset. When a user arrives for the first time or new content is presented, they lack a behavioral history that would allow the recommendation engine to predict their interests. This absence results in the inclusion of new users in broad generic segments, negatively affecting click-through and conversion rates. Without an optimized system, there is a risk that users will drift away before the system has the chance to learn their preferences.
To overcome this issue, a solution has been developed that generates detailed interest profiles from day one. By utilizing large-scale language models, it is possible to synthesize rich representations of users and items without requiring weeks of interaction data. This methodology transforms a cold start into a more responsive and personalized experience.
The designed solution employs Amazon EC2 Trainium chips, optimizing model implementation through deep learning containers (DLC) with the AWS Neuron SDK. This approach enables machine learning engineers to test different configurations of language models and encoders, facilitating rapid iterations on recommendation metrics without needing to modify the base model’s code.
For the development of this innovation, the Amazon book review dataset was used, which provides reviews from real users and metadata for thousands of books. This resource allows for the simulation of “cold start” situations where a new user begins with only one initial review. By implementing a language model that enriches the user’s profile, related subtopics that they might enjoy can be inferred.
The next step in the process involves converting both the expanded interests and the book catalog into comparable vectors using encoders such as Google T5. The impact of different sizes of these encoders on the quality of matches is analyzed, conducting fast searches through FAISS indices.
In later instances, the quality of the recommendations is measured, revealing that as the size of the models increases, the generated signals become more discriminative. This analysis allows developers to find the optimal combination of models and encoders that maximizes performance without incurring high costs.
Looking ahead, there are plans to implement these models in a production environment, ensuring that enriched user profiles are ready to connect with a larger volume of content. This initiative not only demonstrates the transformative potential of machine learning technology in recommendation systems but also promises to enhance the user experience from their very first interaction with the platform.
Referrer: MiMub in Spanish