Feature Selection in the Data Stream Based on Incremental Markov Boundary Learning.

Area of Science:

• Data Mining
• Machine Learning
• Real-time Systems

Background:

• High-dimensional streaming data in real-time systems pose significant hardware and software challenges.
• Existing feature selection algorithms for streaming data often fail due to distribution shifts in nonstationary environments.
• Performance degradation occurs when underlying data distributions change, impacting the reliability of current algorithms.

Purpose of the Study:

• To investigate feature selection in streaming data robust to distribution shifts.
• To propose a novel algorithm based on incremental Markov boundary (MB) learning.
• To enhance the robustness of feature selection in nonstationary data streams.

Main Methods:

• Employs incremental Markov boundary (MB) learning by analyzing conditional dependence/independence.
• Transforms learned information from previous data blocks into prior knowledge for current data blocks.
• Monitors the likelihood of distribution shift and the reliability of conditional independence tests.

Main Results:

• The proposed MB learning approach is naturally more robust against distribution shifts compared to prediction-focused methods.
• Prior knowledge integration assists MB discovery while avoiding negative impacts from invalid information.
• Extensive experiments show the superiority of the proposed algorithm on synthetic and real-world datasets.

Conclusions:

• The novel incremental Markov boundary learning algorithm effectively handles feature selection in streaming data with distribution shifts.
• This approach offers improved robustness and performance in nonstationary scenarios.
• The method provides a more reliable mechanism for feature selection in dynamic data environments.