A neural network model for identification and recovery of dissociated data based on adaptive resonance

Most of the existing machine learning clustering algorithms demonstrate strong performance when applied to complete datasets. However, a noticeable decline in clustering accuracy and robustness is often observed when data are sparse or contain missing attributes. In this research, two modifications of the ART-2 adaptive resonance network are proposed, namely ART-WD and ART-DWD, designed for the classification of sparse data with missing attributes. The concept of “weakly defined data” is introduced to characterize data that can only be obtained in limited amounts, infrequently, and irregularly, as well as relationships lacking a clear definition. ART-WD is designed to classify weakly defined vectors and infer their potential uniqueness, while ART-DWD is derived from ART-WD to facilitate the recovery of missing attributes. The developed models were evaluated on recognition and recovery tasks on three datasets, including the children’s blood zinc content dataset, the iris dataset, and the vehicle exhaust lead pollution dataset. Additionally, comparative computational experiments were conducted on the five most popular classification methods, demonstrating the high efficiency of the developed approaches. Code is available athttps://github.com/DABINB/ART-DWD.