Concept drift detection in toxicology datasets using discriminative subgraph-based drift detector.

Area of Science:

• Computational toxicology
• Graph stream analysis
• Machine learning

Background:

• Graphs and graph streams are increasingly vital for data representation.
• Concept drift detection in graph streams is under-researched, especially in toxicology.
• Existing methods lack application in toxicological graph stream analysis.

Purpose of the Study:

• To apply the discriminative subgraph-based drift detector (DSDD) to graph streams from toxicology datasets.
• To evaluate DSDD's performance with different heuristics (MDL and SIZE) and window sizes.
• To assess the impact of DSDD on graph stream classification using a long short-term memory model.

Main Methods:

• Generated graph streams from four toxicology datasets, creating abrupt and gradual drift scenarios.
• Applied DSDD with Minimum Description Length (MDL) and SIZE heuristics to detect concept drift.
• Compared DSDD's effectiveness across varying window sizes and drift types.
• Integrated DSDD with a long short-term memory (LSTM) based graph stream classifier.

Main Results:

• DSDD demonstrated effectiveness in detecting concept drift in toxicology graph streams.
• Performance varied based on the heuristic used (MDL vs. SIZE) and window size.
• Drift detection using DSDD improved the performance of the LSTM graph stream classification model.

Conclusions:

• The discriminative subgraph-based drift detector (DSDD) is applicable and effective for concept drift detection in toxicology graph streams.
• The choice of heuristic and window size impacts DSDD's performance.
• Implementing DSDD enhances the accuracy of graph stream classification in toxicology.