Co-Training Semi-Supervised Learning for Fine-Grained Air Quality Analysis

Due to the limited number of air quality monitoring stations, data collected are limited. Using supervised learning for fine-grained analysis, that is used predict index (AQI) locations without may lead overfitting in models have superior performance on training set but perform poorly validation and testing set. In order avoid this problem learning, most effective solution increase amount data, study, not realistic. Fortunately, semi-supervised can obtain knowledge from unlabeled samples, thus solving caused by insufficient samples. Therefore, a co-training method combining K-nearest neighbors (KNN) algorithm deep neural network (DNN) proposed, named KNN-DNN, which makes full use samples improve model analysis. Temperature, humidity, concentrations pollutants source type as input variables, KNN DNN learners. For each learner, labeled initial relationship between variables AQI. iterative process, labeling pseudo-sample with highest confidence selected expand The proposed evaluated real dataset stations 1 February 30 April 2018 over region 118° E–118°53′ E 39°45′ N–39°89′ N. Practical application shows has significant effect analysis quality. coefficient determination predicted value true 0.97, better than other models.