A Research on the Application of Adaptive Rotation Deep Forest on Low Dimensional Datasets
Hongwei Chen, Dewei Shi, Xun Zhou, Fengzhou Zhang · Proceedings of the 2021 5th International Conference on Electronic Information Technology and Computer Engineering · 2021
In recent years, the research on deep learning is based on neural networks and mainly trains multilayer nonlinear parameters through a back propagation algorithm. Recent research on deep forest represented by gcForest opens a door to explore non-neural network depth models. This model is a deep learning model based on random forest, a training process of which does not depend on the backpropagation algorithm. Compared with the general ensemble learning method, it has better classification accuracy, especially on small-scale training sets. Low-dimensional data is also very common in the digital age. Because of the small number of features available for learning, it is very challenging for both neural networks and general ensemble learning models. This paper proposes a rotated adaptive deep forest (RADF) model based on improved RotBoost. The model has two main characteristics: Firstly, RotBoost is introduced to solve the problem that data cannot be scanned with multi-granularity due to a lack of spatial correlation to improve the diversity of low-dimensional data. Secondly, for better adapting to low-dimensional data, we use dense connections to optimize the cascade structure to provide more learning features for the model. Experiments show that the performance of RADF on low-dimensional data sets is better than that of deep forest and some mainstream ensemble learning models.