Based on Clustering Label Generation for View Type Noisy Learning

Based on the previously designed View type detection system, there is already a comprehensive algorithm framework for achieving complete View type detection functionality. Due to the increasing variety and quantity of apps nowadays, as well as the different scenarios contained in each app, we want to collect more data for model training and testing. However, the cost of using manual annotation increases with the number of views that need to be detected. This paper implements a clustering based automatic generation of labels to achieve semi supervised learning. However, considering that the automatically generated labels may contain incorrect labels (which are inconsistent with the manually annotated results), Co-teaching method was adopted to perform noisy learning on the data containing incorrect labels. To reduce the error rate of label generation, we improved the DPC-MND algorithm by introducing a dynamic truncation threshold and improving its allocation strategy, thereby enhancing the algorithm's ability to handle noisy or irrelevant points. And established a mapping method for clustering centers to labels, achieving automated generation of labels.