BIM command recommendation using dynamic graph neural network

Sequential recommendation systems predict the next item a user will likely interact with using temporal patterns and contextual information learned from historical user-item interactions. This study applies a Dynamic Graph Neural Network for Sequential Recommendation (DGSR) to predict the next BIM authoring command and extends it to the inductive setting. The proposed method models the user-command interactions in the log data as a dynamic graph and trains a DGSR model to generate embeddings that capture sequential and structural information. A similarity-based weighted average aggregation is introduced for new, unseen users to transfer embeddings from the pre-trained user nodes to the new ones based on their initial interactions. These aggregated embeddings are used to predict the new user's next preference. The model is then partially retrained to enhance the new users' representations. This hybrid approach combines the advantages of pre-trained embeddings with adaptive retraining, enabling the originally transductive DGSR to accommodate the growing number of new users in production environments. Evaluations on a real-world BIM log dataset demonstrate that the proposed model offers better prediction performance compared to the Transformer-based method, showcasing the potential of dynamic graph-based recommendation systems in BIM command recommendation scenarios. The code associated with this paper is available at: https://github.com/saqqa95/DGSR_Induction