Date of Award

10-17-2024

Publication Type

Thesis

Degree Name

M.Sc.

Department

Computer Science

Keywords

Computer Vision;Large Language Models (LLMs);Low-Level Features;Movie Recommended System;Object Detection;Top-N Algorithm

Supervisor

Luis Rueda

Abstract

The abundance of online movie options often overwhelms viewers, making it difficult to choose films that match their preferences. We present a hybrid recommendation framework that employs object detection models, such as YOLOv8, to identify and low-level catalogue objects within movies, creating an item-item similarity matrix of movies. The proposed method involves constructing vectors from the frequency of detected objects within films, which are then used to determine similarity between different movies. We also fetch audio transcripts to extract relevant features using large language models (LLMs). By integrating this information with a Top-N recommendation algorithm, we provide users with suggestions that align with their interests. Our method has the ability to tackle challenges in video recommendation, including the cold-start problem for new or metadata-sparse content. By autonomously analyzing the visual and aural elements of the movies, our system can provide meaningful recommendations, even in the absence of movie metadata, which is a common occurrence on content platforms such as YouTube. With the intersection of computer vision, recommender systems, and LLMs, incorporating low-level features extracted from movies significantly enhances the precision of recommendations, suggesting a promising direction for further developments in recommendation systems. In addition, we plan to make the extracted object dataset available to the public for further research and to promote computational efficiency, which can benefit the environment.

Download

Share

COinS