@

초록

This study presents a multimodal deep learning-based prediction model tailored to the unique transaction dynamics of secondhand trading, aimed at enhancing user experience and transaction success rates on secondhand trading platforms. The model integrates multiple factors central to these transactions - including product images, post descriptions, seller transaction history, and the number of favourites - through an early fusion approach that combines InceptionResNet (Inception-Residual Network), RoBERTa (Robustly Optimised Bidirectional Encoder Representations from Transformers), BiGRU(Bidirectional Gated Recurrent Unit), and MLP (Multilayer Perceptron) architectures. Achieving an accuracy of 79.2%, the model identifies seller trust as the most influential factor in transaction success, followed closely by clear, positive post descriptions and competitive pricing. Sellers with higher trust scores, extensive transaction experience, and well-set prices demonstrated higher transaction completion rates. In contrast, product images and the number of favourites had comparatively less impact on outcomes. These findings carry practical implications for platform operators and sellers. Platforms should implement a structured framework to improve seller trust scores and offer transparent pricing guidelines to enhance transaction success rates. Additionally, detailed, trustworthy product descriptions were shown to play a key role in driving successful transactions, underscoring the importance of comprehensive and buyer-focussed listing practices.

키워드