Label Distribution Learning Method Based on Deep Forest

Label Distribution Learning (LDL), as a novel paradigm aimed at resolving label ambiguity, has attracted significant attention in recent years. To further improve the predictive accuracy of LDL, we propose a LDL Method Based on Deep Forest (DFLDL). This method capitalizes on the cascade structure of deep forests, processing input data layer by layer to efficiently capture and model complex feature representations. To mitigate potential degradation in deep models, DFLDL integrates a confidence function that ensures the maximal exploitation of features at each layer, thereby minimizing training errors and enhancing predictive outcomes. Experimental results reveal that DFLDL surpasses other LDL methods across various datasets, demonstrating enhanced capabilities.