Abstract:Gradient boosting decision tree algorithm is widely used in various tasks, such as classification, regression, and ranking, owing to its high accuracy and strong interpretability. With the explosive growth of data volume, distributed gradient boosting decision tree algorithms have become an important research issue. Although there exists a series of implementations of distributed gradient boosting decision tree, they perform poorly on high-dimensional and multi-classification tasks. The data parallel strategy they adopt requires the transmission of gradient histograms, and this communication overhead becomes the bottleneck in many high-dimensional and multi-classification task. This study aims at this problem and tries to find an efficient parallel strategy that is more suitable for the target. Data-parallel and feature-parallel strategies are first compared based on a cost model, and it is theoretically proved that feature-parallel is more suitable for high-dimensional and multi-classification tasks. Based on the analysis, this paper proposes a feature-parallel distributed gradient boosting decision tree algorithm, named FP-GBDT. FP-GBDT designs an efficient distributed dataset transposition method to partition the training dataset by column. During the construction of gradient histogram, FP-GBDT uses a sparsity-aware method to accelerate the histogram construction. When splitting tree nodes, FP-GBDT develops a bitmap compression method to transmit the placement of instances, thereby reduces the communication overhead. This study compares the performance of distributed gradient boosting decision tree algorithm under different parallel strategies through extensive experiments. First, the effectiveness of proposed optimization methods in FP-GBDT is verified. Then, the representative of data-parallel strategy of FP-GBDT and XGBoost are compared. On various datasets, it is proved that FP-GBDT is more efficient in high-dimensional and multi-classification tasks. FP-GBDT achieves up to 6 times performance improvement than data-parallel implementations.