Genetic Algorithm-Based Two-Step Optimization for Precise Latent Factor Analysis.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Latent Factor Analysis (LFA) is crucial for extracting information from high-dimensional and sparse (HiDS) matrices.
• Traditional LFA models trained with Stochastic Gradient Descent (SGD) face accuracy limitations due to manual learning rate adjustment and reliance solely on gradient information.
• Particle Swarm Optimization (PSO) offers adaptive learning rates but struggles with dynamic decision spaces due to strong convergence.

Purpose of the Study:

• To propose a novel LFA model, Genetic Algorithm-based Two-Step LFA (GA-TSLFA), to overcome the limitations of traditional LFA training methods.
• To leverage the flexibility of Genetic Algorithms (GA) for hyperparameter tuning in dynamic decision spaces and refining models in complex, high-dimensional spaces.
• To improve the prediction accuracy and training efficiency of LFA models for HiDS matrices.

Main Methods:

• GA-TSLFA employs a two-step training process.
• Step 1: Pre-trains the LFA model using SGD with a learning rate adaptively adjusted by a proposed GA.
• Step 2: Refines the latent factor (LF) matrices using a GA-based framework that optimizes selected partial vectors via a dedicated strategy.

Main Results:

• Empirical studies on benchmark datasets demonstrate that GA-TSLFA achieves superior prediction accuracy compared to state-of-the-art LFA models.
• The proposed GA-TSLFA model exhibits competitive training efficiency.
• The GA-based optimization framework effectively refines LF matrices, enhancing overall model performance.

Conclusions:

• GA-TSLFA offers a significant advancement in LFA model training, particularly for HiDS matrices.
• The integration of GA provides superior adaptability and optimization capabilities compared to PSO for LFA hyperparameter tuning.
• The two-step training approach effectively enhances model accuracy and efficiency, establishing GA-TSLFA as a leading method in the field.