Air Quality Inference with Deep Convolutional Conditional Random Field

Abstract

Conditional Random Field is a discriminative model for time series data. In this paper, we propose an improved CRF and apply it to the task of air quality inference. Different from the classical CRF, our linear chain CRF is a supervised learning based on the deep convolution neural network, which has a strong learning ability and fast processing speed for the engineering big data. Specifically, we model the state feature function and the state transition feature function with deep convolutional neural network. The parameter space can store more feature expressions learned from a large number of data. For the state transition feature function of linear conditional random field, we add the influence of input sequence on this function. Through the modelling and learning both vertex features and edge features from data, we obtain a more powerful and more efficient CRF. Experiments on natural language and air quality data show our CRF can achieve higher accuracy.