Online Causal Feature Selection for Streaming Features.

Area of Science:

• Machine Learning
• Causal Inference
• Data Mining

Background:

• Causal Feature Selection (CFS) offers superior interpretability and predictability.
• Existing methods like max-min Markov Blanket (MB) discovery are not suited for streaming data.
• Online Streaming Feature Selection (OSFS) and Granger Selection Method (GSM) have limitations in MB discovery and prediction accuracy for streaming data.

Purpose of the Study:

• To propose an Online Causal Feature Selection for Streaming Features (OCFSSFs) algorithm.
• To address limitations of existing methods in handling streaming data for accurate Markov Blanket discovery.
• To enhance prediction accuracy by mining the complete Markov Blanket, including parents, children, and spouses.

Main Methods:

• Developed an OCFSSFs algorithm for real-time Markov Blanket discovery in streaming feature environments.
• Employed an interleaving learning method for identifying parents, children, and spouses within the Markov Blanket.
• Distinguished between parents/children and spouses dynamically during the online learning process.

Main Results:

• Experimental evaluation on synthetic datasets demonstrated high precision, recall, and minimal distance.
• Testing on real-world and time series datasets confirmed the algorithm's effectiveness in classification precision and feature selection efficiency.
• The proposed OCFSSFs algorithm showed significant improvements over existing methods in handling streaming data.

Conclusions:

• The OCFSSFs algorithm effectively mines the complete Markov Blanket for streaming features in real time.
• The method enhances prediction accuracy by accurately identifying all relevant causal relationships, including spouses.
• OCFSSFs offers a robust solution for causal feature selection in dynamic and high-velocity data streams.