Sparse Time Series Representation in Public Events Early Warning Model

Abstract

To solve the problem of mechanisms’ delay in early warning for unexpected public security incidents, this paper proposed a highly effective time series sparse representation for determining early warning for public incidents on the basis of position and reinforcement learning. Fractal and sparse were introduced to denote multi-dimensional time series, which effectively reduced the dimensions of time series while avoiding their distortion. Reinforcement learning was used to build the model of early warning for public incidents. Paths to anomaly series or early warning for public incidents were obtained by the agent through iteration training. Then weight parameters of neural networks were analyzed for the determination of early warning for incidents. Simulation showed that under this algorithm, convergence happened when the number of steps was in the range from 500 to 800, 37.5% smaller than that when using the original data. This result of the experiment demonstrated that this method greatly improved the efficiency of early warning for public incidents.