New Splitting Criteria for Decision Trees in Stationary Data Streams

Area of Science:

• Data Mining
• Machine Learning
• Statistical Learning Theory

Background:

• Hoeffding decision trees are popular for stream data mining but are heuristic.
• Existing methods based on Hoeffding's inequality have limitations for classical decision trees like ID3 or CART.
• There is a need for mathematically sound and high-performing algorithms for stream data mining.

Purpose of the Study:

• Develop new, mathematically justified splitting criteria for classification in stationary data streams.
• Investigate the probabilistic properties of these novel criteria.
• Address the limitations of heuristic approaches in stream data mining.

Main Methods:

• Developed a family of new splitting criteria for stationary data streams.
• Derived criteria based on misclassification error and Gini index impurity measures.
• Proposed a general division of splitting criteria into two types and introduced hybrid criteria.

Main Results:

• New splitting criteria offer probabilistic guarantees on attribute selection.
• Type- criteria ensure high probability of selecting the optimal attribute for the entire data stream.
• Hybrid criteria combine misclassification error and Gini index for improved performance.

Conclusions:

• The developed splitting criteria provide a mathematically rigorous foundation for stream data mining.
• These criteria offer improved performance and reliability compared to heuristic methods.
• The proposed framework enhances classification accuracy in stationary data streams.