LogPar: Logistic PARAFAC2 Factorization for Temporal Binary Data with Missing Values.

Area of Science:

• Data Science
• Machine Learning
• Tensor Decomposition

Background:

• Real-world binary data often contains missing values, represented as irregular tensors.
• Existing tensor factorization models fail with missing values and assume incorrect distributions for binary data.

Purpose of the Study:

• To develop a novel model, Logistic PARAFAC2 (LogPar), for accurate low-rank approximation of binary irregular tensors with missing values.
• To address the distribution mismatch and improve performance in tensor completion and downstream tasks.

Main Methods:

• Modeled binary irregular tensors using a Bernoulli distribution parameterized by an underlying real-valued tensor.
• Employed a positive-unlabeled learning loss function to handle missing values.
• Incorporated uniqueness and temporal smoothness regularization for enhanced interpretability.

Main Results:

• LogPar achieved up to 26% relative improvement in irregular tensor completion compared to existing baselines.
• Demonstrated significant performance gains in downstream predictive tasks, including 13.2% for heart failure and 14% for mortality prediction.

Conclusions:

• LogPar effectively handles binary irregular tensors with missing values, outperforming state-of-the-art methods.
• The model offers improved accuracy and interpretability for complex, real-world datasets.