Dynamic Feature Acquisition Using Denoising Autoencoders.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Real-world machine learning models face varying feature acquisition costs during testing.
• Optimizing the trade-off between performance and cost is crucial for practical applications.
• Existing methods may not adequately address the dynamic nature of feature costs and context.

Purpose of the Study:

• To introduce a novel, scalable, and cost-aware method for feature acquisition at test time.
• To develop a system that incrementally requests features based on context and acquisition cost.
• To optimize the performance-cost balance in machine learning models operating in real-world scenarios.

Main Methods:

• Utilizing sensitivity analysis in neural networks to measure feature informativeness.
• Employing denoising autoencoders with binary representation layers for density estimation.
• Integrating a context-dependent informativeness measure based on prediction sensitivity to unknown features.

Main Results:

• The proposed method demonstrated efficient feature acquisition in a cost- and context-aware manner.
• Evaluated on eight real-world and one synthesized dataset, showing superior or competitive performance.
• Successfully handled unknown features using denoising autoencoders.

Conclusions:

• The developed approach offers an effective solution for cost-aware feature acquisition in practical machine learning.
• The method's scalability and adaptability to context-dependent feature costs are key advantages.
• This work advances the field of efficient model deployment by optimizing resource utilization at test time.