Anticipative Bayesian classification for data streams with verification latency.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Adaptive classification in non-stationary data streams often requires recent labeled data, which is frequently unavailable.
• Existing methods for stream classification with verification latency have limitations, such as assuming clustered data or homogeneous feature drift.

Purpose of the Study:

• To propose the Anticipative Bayesian stream Classifier (ABClass), an adaptive classification approach for non-stationary data streams.
• To address the challenge of missing labels and limited applicability of current methods by enabling integration and automatic selection of components.

Main Methods:

• ABClass employs a Bayesian classification framework combining density estimation with extrapolation for drift patterns.
• Unsupervised parameter tuning and model selection are utilized, allowing for multivariate density estimation and extrapolation.
• Feature-specific drift patterns are modeled assuming conditional independence between features given the class label.

Main Results:

• ABClass demonstrates competitiveness against state-of-the-art approaches on real-world data streams.
• The proposed method shows significant speed improvements, being ten- to hundred-times faster than competitors for both model fitting and prediction.
• ABClass is generative, facilitating the explanation and visualization of concept drift patterns.

Conclusions:

• ABClass offers an effective solution for adaptive stream classification, particularly in scenarios with missing labels and verification latency.
• Its generic nature allows for the integration of diverse drift models, enhancing its adaptability.
• The computational efficiency and performance gains make ABClass a practical choice for real-time data stream analysis.