Non-Gradient Hash Factor Learning for High-Dimensional and Incomplete Data Representation Learning.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• High-dimensional and incomplete (HDI) data are prevalent in Big Data applications like drug discovery and recommender systems.
• Hash learning offers efficient representation learning for HDI data due to fast reasoning and low storage requirements.
• Existing hash learning methods suffer accuracy loss from quantization during gradient-based optimization of discrete hash factors.

Purpose of the Study:

• To propose a novel non-gradient hash factor (NGHF) model to address the accuracy limitations of current hash learning techniques for HDI data.
• To develop an effective discrete optimization strategy that avoids quantization loss inherent in gradient-based methods.
• To enable precise binary representation of HDI data with high learning ability.

Main Methods:

• Introduced a discrete differential evolution (DDE) algorithm to simulate continuous optimization for binary codes, optimizing the discrete learning objective directly.
• Applied the DDE algorithm to the NGHF model for efficient and precise training without quantization loss.
• Provided theoretical convergence guarantees for the NGHF model.

Main Results:

• The NGHF model demonstrated high representation learning ability, comparable to real-valued models for HDI data.
• Extensive experiments on nine real-world datasets showed NGHF significantly outperformed eight state-of-the-art hash learning models.
• NGHF achieved accuracy comparable to real-valued models in HDI data representation learning.

Conclusions:

• The NGHF model effectively overcomes quantization loss, achieving high accuracy and fast reasoning for HDI data representation.
• The proposed non-gradient approach enhances hash learning models for critical industrial Big Data applications.
• NGHF offers a promising solution for precise binary representation of complex datasets.