Utilização de filtragem colaborativa no auxílio de recomendação personalizada para leitores de mangá

Abstract

This study investigated, developed, and compared two approaches for generating manga recommendations: the Singular Value Decomposition (SVD) model and the Pearson Correlation Coefficient. The methodology involved data preparation through the development and execution of a web scraper to extract manga information and reviews from a highly active internet forum. Challenges arising in the applicability of these data extraction methods were addressed, along with alternatives for handling source blocking situations, model training, and performance evaluation, focusing on collaborative filtering and personalized recommendations for user profiles and manga works. In the implementation of SVD, latent patterns in user review data were identified, enabling personalized recommendations based on individual preferences through the sharing of experiences with similar profiles. However, metrics such as Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) revealed the need for model refinement to improve its accuracy, as well as alternative implementations for conducting comparisons and metrics related to the specific data mass used in the study. Conversely, the approach based on the Pearson Correlation Coefficient prioritized similarity between manga reviews to generate item-focused recommendations, although it significantly relied on the number of available reviews. This methodology offered a direct and valid logic for personalized recommendations based on relationships derived from reviews. The conclusion highlighted the future possibility of exploring hybrid methods combining the advantages of SVD and the Pearson Correlation Coefficient to achieve more precise and comprehensive recommendations, as well as validating techniques that bring different recommendation approaches for tangible comparison. The utilization of additional data gathered in the generated data mass to enrich the quality of recommendations was suggested, aiming to use more detailed parameters in recommendations, along with the employment of indirect approaches, such as using LLMs to aid in the recommendation process. Finally, the study emphasizes the importance of advancing these recommendation technologies to facilitate readers' lives by assisting in filtering the vast content offered by the industry and the internet.