Deep multimodal fusion for video game age rating classification

Video games appeal to a wide range of ages, from children to adults. As a result, reliable age rating systems like the Entertainment Software Rating Board (ESRB) and Pan European Game Information (PEGI) are essential for guarding younger gamers from improper content. These organizations rate games based on content submitted by video game developers. This paper proposes a multimodal deep learning framework that predicts age ratings by analyzing both video game cover images and textual descriptions. A dataset of 39,212 games was constructed using publicly available information, including ESRB and PEGI labels. Both individual models based on visual or textual features and fusion models that combine these modalities using simple concatenation and Deep Canonical Correlation Analysis (DCCA) were employed to perform the classification task. Experimental results indicate that the simple concatenation model achieves the highest accuracy compared to the individual modalities and the DCCA-based approach, reaching 0.678 for ESRB prediction and 0.584 for PEGI prediction. The findings highlight that using only visual information has limitations, and that textual descriptions play an important role in determining the appropriate age rating for a game. This study shows that future research can benefit from using additional content like gameplay videos and audio.