Development of a multimodal video game recommendation system

This study presents the development and evaluation of a multimodal video game recommendation system utilizing a dataset from the Steam platform. The system incorporates various data types, such as game descriptions, screenshots, and categories, to improve recommendation accuracy. Two dataset versions were created: one based solely on the basic characteristics of the games, and another enhanced with probabilities derived from machine learning models used for text sentiment analysis, screenshot content identification, and object detection. These datasets were processed using sparse matrices, standardized, and reduced in dimensionality through TruncatedSVD. The Annoy library was employed for indexing, and a web application was built on the Streamlit platform for user testing. The findings indicate that the model augmented with machine learning models significantly outperforms the basic version, highlighting the effectiveness of integrating multimodal data.