Loading...

Design and Implement Large Language Models as a Recommender Systems

Ghayouri Sales, Ali Akbar | 2024

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 57361 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Motahari, Abolfazl; Abaam, Mohammad Ali
  7. Abstract:
  8. Advancements in the field of Natural Language Processing have led to the evolution of recommender systems based on natural language processing. However, current models predominantly perceive products as identifiers without any inherent information during modeling, mapping each product to a unique identifier. This approach creates several significant limitations:
    • There is no full utilization of the content information of the products and the lexical capacities of natural language processing models.
    • Analyzing user interests to obtain relevant and high variance products is not possible.
    • Adapting to practical scenarios such as product availability in stores and adding product information to the recommender system space is challenging and sometimes impossible. In an effort to address these limitations, we introduce an innovative and generative framework, inspired by the integration of Large Language Models with Semantic Search Engines. The primary concept revolves around using a large language model as a functional unit in understanding user queries and assisting users in generating optimal queries for the semantic search engine. This framework reduces the uncertainty of users about their exact needs by allowing editing and interaction with the recommender system. It utilizes a semantic search engine to provide the most suitable output for purchasing products available on the system. Moreover, the queries generated in this system serve as interpretable questions for the user. This is particularly useful for providing recommendations for products with minimal or no prior interaction, a problem known as the "cold start" issue in recommender systems. Consequently, in this framework, we have a search and recommendation system that is interactive, interpretable, and generative. In this research, this product is implemented in the apparel space as a web-based software. The reason for choosing the apparel space is the presence of a large, complex, diverse dataset, and the necessity of interacting with the user to obtain the ideal product in this area.
  9. Keywords:
  10. Recommender System ; Large Language Model ; Natural Language Processing ; Machine Learning Software

 Digital Object List

 Bookmark

...see more