Idea
An unsupervised multi-view feature selection platform that handles incomplete data for improved data analysis and machine learning.
Research Paper
Core Innovation
This paper introduces TRUST-FS, which uniquely combines feature selection, missing data imputation, and view weight learning in a single tensor factorization model. It addresses missing variables within views rather than just missing views, and uses Subjective Logic to create a trustworthy similarity graph that guides the entire process. This integrated approach improves accuracy and robustness compared to prior methods that treat imputation and feature selection separately.
Market Size (TAM)
$2–10B TAM for multi-view data analytics and feature selection tools; $1–3B SAM from enterprises and research institutions handling incomplete multi-source data. Driven by growing adoption of AI in heterogeneous data environments and demand for robust unsupervised learning.
Potential Customers & Pain Points
- Data Scientists handling multi-view incomplete datasets
- Machine Learning Engineers needing reliable feature selection
- Enterprises with heterogeneous data sources suffering from missing data
- Researchers working on unsupervised learning with incomplete features
Business Model
Offer TRUST-FS as a SaaS platform or API for data preprocessing and feature selection; provide enterprise licensing and consulting for integration and customization.
Competitive Landscape
- AutoImpute
- MultiViewFS
- DeepFeatureSelect
Implementation Challenges
- Complexity of tensor factorization for large-scale data
- Integration with existing data pipelines
- Adoption resistance due to novel methodology
Validation Strategy
- Benchmark TRUST-FS against state-of-the-art methods on public multi-view datasets
- Pilot deployments with enterprise clients handling incomplete multi-source data
- Collect user feedback to refine usability and integration features
Research Paper Overview
TRUST-FS: Tensorized Reliable Unsupervised Multi-View Feature Selection for Incomplete Data
Summary
This paper proposes TRUST-FS, a novel multi-view unsupervised feature selection method that handles incomplete multi-view data with missing variables. It integrates feature selection, missing-variable imputation, and view weight learning into a unified tensor factorization framework using adaptive-weighted CP decomposition. TRUST-FS leverages Subjective Logic to build a reliable similarity graph, improving feature selection and imputation accuracy. Experiments show its superiority over state-of-the-art methods.