What are sequence-based recommendations (SASRec, BERT4Rec)?

Sequence-based recommendations are an approach in recommendation systems that considers the order of user interactions with items (e.g., viewed products, listened tracks) to predict the next interesting object for the user.

SASRec (Self-Attentive Sequential Recommendation) uses the self-attention mechanism from transformers to model the sequence of user actions. This allows for effective consideration of long-term dependencies and context, improving recommendation quality.

BERT4Rec is an adaptation of the BERT model for recommendations. It is trained to predict missing elements in the interaction sequence using bidirectional attention, which helps better understand the user's context and preferences.

Both methods outperform traditional models that only consider the last actions or simple statistics, thanks to deep analysis of sequences and context.