Deep direct likelihood knockoffs.

Area of Science:

• Machine Learning
• Computational Biology
• Statistical Modeling

Background:

• Black box machine learning models, like deep neural networks, excel at prediction but obscure feature importance.
• Identifying key predictive features is crucial for scientific discovery and guiding costly experiments.
• Existing methods like Model-X knockoffs require complex generative models for valid feature selection.

Purpose of the Study:

• To introduce Deep Direct Likelihood Knockoffs (ddlk), a new method for feature selection in predictive modeling.
• To enable reliable discovery of important features while controlling the false discovery rate (fdr).
• To overcome limitations of existing knockoff methods by directly optimizing the knockoff property.

Main Methods:

• Developed ddlk, a two-stage method that maximizes feature likelihood and minimizes KL divergence for knockoff validity.
• Utilized the Gumbel-Softmax trick to optimize the knockoff generator against worst-case swaps.
• Validated ddlk on synthetic datasets and real-world benchmarks, including COVID-19 data.

Main Results:

• ddlk demonstrated superior power in feature discovery compared to baseline methods.
• The method effectively controlled the false discovery rate across various benchmarks.
• Achieved state-of-the-art performance in identifying significant features.

Conclusions:

• ddlk provides a robust and powerful approach for feature selection in machine learning, particularly in scientific domains.
• The method enhances the reliability of feature discovery by controlling the false discovery rate.
• ddlk is applicable to diverse datasets and predictive modeling tasks.