ContrastVAE++: 사후 붕괴 완화와 언어모델 기반 데이터 증강을 통한 대조적 변분 오토인코더의 성능 향상

Sequential recommendation systems utilize the user's temporal behavioral information to identify preferences and provide relevant items. ContrastVAE, a VAE-based SOTA recommendation model, has proposed an effective method for sequential recommendation by combining contrastive learning and data augmentation. However, ContrastVAE has a posterior collapse problem in which the decoder does not consider the encoder condition when generating sequential information, and it uses a data augmentation technique inefficient to learn the latent space. To overcome these limitations, this study proposes ContrastVAE++, which alleviates the posterior collapse problem and applies an advanced augmentation technique. ContrastVAE++ employs context-based data augmentation to consider the correlation between items, Dirichlet-distributed noise to apply a distribution dependent on the encoder output, and decoder cross-attention for effective latent vector representation. Our experimental results show that ContrastVAE++ outperforms ContrastVAE on all experimental data.