Iterative Model Based on Multi-Model Spatio-Temporal Evolution

• 1、State Key Laboratory of Networking and Switching Technology, School of Computer Science，Beijing University of Posts and Telecommunications, Beijing,100876, China

Abstract：In order to achieve informatioIterative Model Based on Multi-Model Spatio-Temporal Evolutionn fusion and comprehensive decision-making for multi-modal data under temporal and spatial evolution, ensure that the intrusion of newly added modal data into the original model is minimized. This paper proposes an iterative model based on temporal and spatial evolution to realize multi-source, multi-modal incremental data modeling and dynamic fusion. This model mainly includes three parts: modal representation and feature extraction, model fusion, and comprehensive decision-making. Modal representation and feature extraction complete the information extraction and compression of modal data and construct a new streamlined feature representation. Model fusion mainly integrates the incremental model and the current model, completes the cross-sharing of data and information, and supports the comprehensive decision-making model. We have designed three model fusion methods for the local and global model fusion levels: Multi-Modal Joint, Cycle Fusion and Steady-State Fusion. These three fusion methods can effectively solve the problem of model structure balance after iteration and ensure the consistency of the path length from each modal data of the model to the output to the greatest extent. The comprehensive decision-making model is mainly based on the comprehensive feature vector to complete the decision-making task. Through experimental tests, the comprehensive decision-making algorithm selects the Broad Learning System algorithm with excellent characteristics such as training time and good classification effect. Using UCI HAR data, experiments verified the feasibility of the iterative model based on multi-modal temporal and spatial evolution. As the iterative process progresses, the model can be updated quickly, and the comprehensive decision-making ability is steadily improved. The iterative model based on HAR data set is compared with the accuracy of single-model prediction increased by 20.76% on average.

Keywords： Iterative model Multi-Modal Model Fusion BLS Neural Network