Bias Aware Probabilistic Boolean Matrix Factorization.

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Boolean matrix factorization (BMF) is crucial for recommendation systems and dimensionality reduction.
• Current BMF methods often assume homoscedastic noise, treating all data points equally.
• Real-world data exhibits diverse stochastic noises, making uniform data distribution assumptions suboptimal.

Purpose of the Study:

• To introduce a novel probabilistic BMF model that accounts for object- and feature-wise biases.
• To address the limitations of existing BMF methods in handling heteroscedastic noise in binary data.
• To develop a Bias Aware BMF (BABF) approach for more accurate data decomposition.

Main Methods:

• Developed a probabilistic BMF model (BABF) incorporating object- and feature-wise bias distributions.
• Evaluated BABF on diverse datasets with varying noise levels, bias, and signal pattern sizes.
• Compared BABF's performance against state-of-the-art factorization methods.

Main Results:

• BABF demonstrated superior accuracy and efficiency in recovering original datasets compared to existing methods.
• The inferred bias levels by BABF showed a highly significant correlation with true biases in both simulated and real-world data.
• BABF effectively handles datasets with different levels of background noise and bias.

Conclusions:

• BABF is the first approach for Boolean decomposition that considers feature-wise and object-wise biases in binary data.
• The proposed model offers a more robust and accurate alternative to traditional BMF methods.
• BABF provides a significant advancement in analyzing noisy binary datasets across various applications.