New Scalable and Efficient Online Pairwise Learning Algorithm

Area of Science:

• Machine Learning
• Data Science
• Computer Science

Background:

• Online algorithms are crucial for processing streaming data and large-scale pairwise learning.
• Existing methods struggle with scalability and efficiency for high-dimensional data due to singly stochastic gradients.

Purpose of the Study:

• To propose a novel dynamic doubly stochastic gradient (D2SG) algorithm for efficient and scalable online pairwise learning.
• To address the limitations of existing algorithms in handling large-scale, high-dimensional datasets.

Main Methods:

• Developed a dynamic doubly stochastic gradient (D2SG) algorithm specifically for online pairwise learning.
• Analyzed the time and space complexity for incorporating new samples, achieving O(d) complexity where d is data dimensionality.
• Provided rigorous theoretical analysis to guarantee statistical accuracy under standard assumptions.

Main Results:

• The D2SG algorithm demonstrates significantly improved speed and scalability compared to existing online pairwise learning methods.
• Experimental results on real-world datasets validate the theoretical findings of the D2SG algorithm.
• The D2SG algorithm shows superior efficiency and scalability for large-scale, high-dimensional data.

Conclusions:

• The proposed D2SG algorithm effectively overcomes the scalability and efficiency challenges in online pairwise learning for high-dimensional data.
• D2SG offers a promising solution for real-world applications involving large-scale streaming data.
• The algorithm achieves a favorable balance between computational efficiency and statistical accuracy.