A novel multi-label deep forest algorithm based on elite preservation strategy and multi-layer feature fusion

Multi-label classification is a popular research direction in the field of machine learning and pattern recognition, and has shown significant application potential in real-life scenarios. However, traditional multi-label classification algorithms still suffer from the low classification accuracy and instability due to the lack of feature diversity. To tackle this issue, this paper investigates the effect of feature diversity on the ensemble model based on the deep forest framework, and subsequently presents a multi-label deep forest algorithm based on elite preservation strategy and multi-layer feature fusion (EMDF). Firstly, the elite preservation strategy is introduced to screen the forests in the cascades layer by layer for reestablishing the cascades structure. This enables the screened cascades to acquire better predicted feature vectors. Secondly, combined with the predicted feature vectors of screened cascades, the multi-layer feature fusion strategy is put forward to enhance the label information of input features and improve the predictive performance of the model. Finally, EMDF is extensively tested on 10 different types of comparative algorithms and 12 various fields datasets. Experimental results show that EMDF achieves better accuracy on multiple metrics such as Hamming loss and Coverage on most datasets. Furthermore, a comprehensive evaluation of algorithm rankings on all metrics also demonstrates the superior stability of EMDF.