A Study on a Webtoon Recommendation System With Autoencoder and Domain-specific Finetuing Bert with Synthetic Data Considering Art Style and Synopsis
Beom-Hun Kim, Cheol-seung Lee, Sang Kil Lim · International Journal on Advanced Science Engineering and Information Technology · 2024
As the global webtoon market experiences rapid and substantial growth, webtoon platforms increasingly recognize the critical need for customized recommendation services that effectively utilize advanced user personalization technology. This strategic approach is essential for strengthening their market competitiveness in an increasingly crowded digital landscape. Conventional recommendation systems typically rely on content-based or collaborative filtering techniques, primarily measuring user similarities. However, these traditional methods often lead to significant challenges, such as the cold start problem and difficulties with long-tail deployment, primarily due to insufficient user data, especially for new or niche content. We propose an innovative recommendation algorithm based on a sophisticated autoencoder architecture to address and overcome these persistent challenges. Our approach involves a comprehensive evaluation of this algorithm alongside specialized art style and synopsis analysis algorithms. This multi-faceted algorithm is designed to extract hidden features from various components of webtoons and utilizes advanced clustering techniques through in-depth similarity analysis. This process enables the system to precisely determine intricate connections between individual webtoons. Furthermore, we implement a specialized autoencoder for style feature extraction in the art style analysis component to enhance and refine our approach. Complementing this, we employ a domain-specific BERT-based model, augmented with extensive data augmentation techniques, for comprehensive synopsis similarity analysis. In this study, the strategic use of autoencoders allows for the efficient and accurate reconstruction of important features from both the art style and synopsis of webtoons. This innovative approach results in a significantly more robust, scalable, and effective recommendation system, capable of handling the diverse and evolving nature of webtoon content.