CUDT: a CUDA based decision tree algorithm.

Area of Science:

• Computer Science
• Data Mining
• High-Performance Computing

Background:

• Traditional decision tree algorithms face latency issues with massive datasets from ubiquitous sensing.
• Existing distributed systems struggle to efficiently process big data generated in real-time.

Purpose of the Study:

• To design and implement a novel parallelized decision tree algorithm optimized for Graphics Processing Units (GPUs).
• To significantly improve data processing latency in big data mining applications.

Main Methods:

• Developed a parallelized decision tree algorithm leveraging NVIDIA's Compute Unified Device Architecture (CUDA).
• Implemented a hybrid system where CPUs manage control flow and GPUs handle intensive computations.
• Conducted performance evaluations comparing the CUDA-based Decision Tree (CUDT) against traditional CPU-based methods like Weka-j48 and SPRINT.

Main Results:

• The CUDT system demonstrates substantial performance improvements over existing algorithms.
• Achieved speedups ranging from 5 to 55 times faster than Weka-j48.
• Showcased an 18-fold speedup compared to SPRINT for large datasets.

Conclusions:

• The proposed CUDT algorithm effectively addresses the latency challenges in big data mining.
• GPU acceleration via CUDA offers a viable solution for processing massive datasets from sensing nodes.
• This approach significantly enhances the efficiency of decision tree classification in big data scenarios.