Algorithm for Double-Layer Structure Multi-label Classification with Optimal Sequence Based on Attention Mechanism

A common approach to multi-label classification is to perform problem transformation, whereby a multi-label problem is transformed into one or more single-label problems. Problem transformation considers label correlations by extending the attributes, but ignores the importance of each feature attribute for different classification targets, weakens the sensitivity of the classifier, and reduces the classification accuracy. Attention mechanism is a model that simulates the mechanism of human brain attention. It mainly emphasizes the influence of some crucial inputs on the output by calculating the attention probability distribution, which has a good optimization effect on the traditional model. Based on this, this paper proposes a two-layer chain structure multi-label classification (ATDCC-OS) algorithm, which incorporates the attention mechanism. This algorithm constructs a two-layer multi-label classification model in order to realize the correlation between labels through inter-layer and intra-layer interaction. At the same time, the attention mechanism is introduced to focus selectively on the sample features, identify more important information for the current task, and further improve the classification performance of the algorithm. Furthermore, an optimal sequence selection algorithm (OSS) is proposed, seeking to label the pecking order, solving the problem of reduced classification accuracy caused by randomly selecting the class label sequence to train the binary classifier by the chain classification model. The OSS will be used to optimize the second-layer chain classification model of ATDCC-OS. Comparisons on seven benchmark data sets with related algorithms verify the effectiveness of ATDCC-OS.