Tri-training and MapReduce-based massive data learning

The real-world applications to massive data processing raise two huge challenges to the traditional supervised machine learning. First, the sufficient training examples to ensure the generalization ability become unavailable, since the task of labeling examples by experts is time-consuming and expensive; second, it is impossible to load massive data into memory, and the response time is unacceptable for training and classifying the massive data by a traditional serial mode. In this paper, we gracefully combine the semi-supervised learning with the parallel computing technique to meet these two challenges together. In detail, (1) the co-training style semi-supervised learning named Tri-training is exploited and revised in order to perform learning from the limited amount of labeled data and the large amount of unlabeled data. In particular, the co-training process of Tri-training approach is revised by introducing a data editing operation to remove the newly mislabeled data before it is used to re-train another classifier. Since the data editing can effectively remedy the unavoidable mislabeling problem in co-training style algorithms, the generalization ability of the learned hypothesis can be improved. (2) Furthermore, the learning algorithm for each individual classifier and the data editing operation in the Tri-training are re-formed as the Google’s MapReduce parallel pattern instead of the traditional memory-resident and serial mode. The MapReduce pattern ensures that each co-training iteration and predicting process in Tri-training could be implemented in parallel on the clusters of commodity PCs. The experiment results on UCI datasets show the improvement of our revised Tri-training over the accuracy, even if it performs learning only with the limited amount of labeled data. The real application to detect the small pulmonary nodes in CT chest images also proves that our revised Tri-training algorithm can effectively reduce the false negative rate and the false positive rate. As well as, the application proves the scalability to massive data due to the adaptation to the MapReduce parallel paradigm.