We leverage advanced Graph Neural Networks (GNN) technology to explore the relationships between user preferences and behaviors. Simultaneously, we employ Graph Convolutional Networks (GCN) to capture user-item interactions, thereby enhancing recommendation accuracy. With a Transformer-based attention mechanism, we focus on key information, understand complex user preferences, and deliver personalised recommendations. Through Reinforcement Learning (RL), the recommendation system learns through user interaction, dynamically adjusts strategies, and optimizes long-term recommendation effectiveness.

We build a cross-modal understanding system by integrating Computer Vision and Natural Language Processing (NLP) technologies and extract and align visual semantics by using the CLIP model. For speech information, we combine the Transformer architecture development to achieve end-to-end speech recognition and semantic understanding. In terms of text processing, we complete Information Extraction (IE) tasks such as Named Entity Recognition (NET) and sentiment analysis based on pre-trained models like BERT and ERNIE. Ultimately, we merge features from various models through a multimodal Transformer, providing users with comprehensive multimodal perception services.