PROJECT TITLE :

Semi-supervised Concept Learning by Concept-cognitive Learning and Concept Space

ABSTRACT:

People are able to naturally combine a small amount of labeled data with a large amount of unlabeled data when they make classification decisions, which is also known as semi-supervised learning (SSL) in the field of machine learning. This occurs in human concept learning. In particular, the process of human concept learning is not only a fixed one in human cognition but also has the potential to change gradually in response to changing environments. Despite this, the traditional SSL algorithms need to be redesigned so that they can take into account newly input data. In this regard, concept-cognitive learning may prove to be an advantageous option because it is able to implement dynamic processes by mimicking the cognitive processes of humans. During this time, numerous SSL methods were designed based on the feature vector information of instances, while ignoring concept structural information, which is a very important process in the organization of human knowledge. In the meantime, this information was not taken into consideration. By employing concept spaces, in which knowledge is represented by hierarchical concept structures, a novel SSL method for dynamic SSL has been proposed and given the name semi-supervised concept learning method (S2CL). This method was developed based on the idea described above. In addition, in order to make the most of both the global and the local conceptual information, we propose an extended version of S2CL (which we will refer to as S2CL) as a method for learning concepts. To be more specific, the purpose of this paper is to effectively exploit unlabeled data. To do so, the paper begins by presenting some new related theories for S2CL (or S2CL) that are based on a regular formal decision context. After that, a novel SSL framework is designed, and its corresponding algorithm is developed. Finally, in order to demonstrate the efficacy of our methods, which include concept classification and incremental learning despite the presence of a significant amount of unlabeled data, we carry out some experiments on a variety of datasets and report our findings.