Kindergarten Safety Education Recommendation Based on Multimodal Fusion and ALBERT Lightweight Model

Current mainstream multimodal recommendation models for kindergarten safety education usually have large parameter scales and high computational overhead, which makes them difficult to deploy on resource-constrained edge devices. Similar deployment constraints also exist in engineering systems involving wireless sensing, electromagnetic environment monitoring, and antenna-assisted edge perception, where real-time inference and low resource consumption are required. To balance lightweight design and multimodal information integration, this paper proposes a kindergarten safety education recommendation framework based on multimodal fusion and an ALBERT lightweight model. The framework uses ALBERT-base-v2 with frozen embedding layers and fine-tunes only the last two Transformer encoder layers as the text encoder. A lightweight ResNet-18 and WaveNet are adopted for image and audio feature extraction, respectively. Trimodal semantic alignment is achieved through 128-dimensional linear projection, and a low-overhead gating mechanism is introduced to dynamically integrate modal features. Finally, a dual-tower architecture and Bayesian Personalized Ranking (BPR) loss are used to realize personalized recommendation. Experimental results show that the proposed model achieves an inference latency of 86.67 ms, a model size of 87.3 MB, an average HR@5 of 91.4%, and an average age-appropriateness score of 4.32. The results indicate that the model effectively balances deployment efficiency, multimodal representation capability, and adaptability to children’s development, providing a practical solution for intelligent kindergarten safety education systems and lightweight edge recommendation applications related to electromagnetic-aware sensing environments.