SEEK-VEC: Robust Latent Structure Discovery via Ensemble Topic Modeling.

Area of Science:

• Computational biology
• Data science
• Statistical modeling

Background:

• Topic modeling is crucial for uncovering latent structure in count data.
• Standard methods face limitations like restrictive assumptions, noise sensitivity, and topic number misspecification, especially for non-textual data.

Purpose of the Study:

• Introduce SEEK-VEC (Spectral Ensembling of topic models with Eigenscore for K-agnostic Vocabulary Embedding and Classification), an ensemble framework for count data analysis.
• Develop a method that automatically reinforces signal, mitigates noise, and generates a consensus low-dimensional embedding.
• Enable variable classification, pattern discovery, and model diagnostics through prioritization and grouping scores.

Main Methods:

• Utilizes a spectral ensembling procedure to integrate insights from multiple candidate topic models.
• Applies eigenscore for K-agnostic vocabulary embedding and classification.
• Generates prioritization and grouping scores for data interpretation.

Main Results:

• SEEK-VEC demonstrates robustness in realistic settings and outperforms state-of-the-art methods, particularly with weak signal strength.
• Successfully applied to diverse datasets including psychopathology, food preferences, and single-cell transcriptomics.
• Reveals scientifically meaningful latent structures in real-world data.

Conclusions:

• SEEK-VEC offers a powerful and robust approach for latent structure discovery in count data.
• The framework effectively handles noise and avoids restrictive assumptions of traditional topic models.
• Provides valuable tools for data classification, pattern discovery, and diagnostic analysis across various scientific domains.