Multimodal Sequential Recommendation Window Co-Attention

The system segments the recommendation model into distinct modules: an item sequence encoder that processes content item sequences, an information sequence encoder that processes additional information sequences, and a fusion module that combines them. This segmentation allows each module to specialize in processing specific types of sequence data while maintaining overall system manageability.

The patent introduces a temporal dimension by applying transformer-based self-attention mechanisms that capture dynamics across time steps in sequences. Instead of treating sequences as static inputs, the model processes temporal relationships between items, adding a time-aware dimension to the recommendation approach that improves accuracy without proportionally increasing complexity.

The system applies attention mechanisms selectively rather than uniformly across all sequence elements. By focusing computational resources on capturing critical temporal relationships and interactions between item and information sequences, the model achieves high recommendation accuracy without the full computational overhead of applying complex attention to every possible sequence combination.

The patent merges the item sequence encoder and information sequence encoder into a unified transformer-based architecture that processes both sequences simultaneously. This merging allows the model to capture interactions between items and additional information in an integrated manner, improving recommendation accuracy while sharing computational resources across both encoding tasks.

The system performs preliminary encoding of content item identifiers into dense vector representations before processing. This preliminary action transforms discrete identifiers into continuous embeddings that capture semantic relationships, enabling faster subsequent processing while preserving and enhancing modality information rather than losing it.

The patent introduces transformer encoders as intermediary components that bridge content item identifiers and additional information sequences. These encoders act as mediators that process and integrate multiple modalities of information, preventing information loss while maintaining efficient processing through learned representations.