Global tests for novelty

Area of Science:

• Statistics
• Machine Learning
• Bioinformatics

Background:

• Outlier detection identifies unusual observations, while novelty detection finds exceptional new data points compared to training data.
• Often, the presence of novelty itself is more critical than pinpointing individual novel instances, such as in screening new cancer treatments.

Purpose of the Study:

• To develop and validate hypothesis tests for global level novelty detection.
• To introduce innovative methods applicable under general assumptions, advancing current literature.

Main Methods:

• Development of novel test statistics operating on local neighborhoods.
• Utilizing the permutation principle to derive the null distribution of the test statistics.
• Assessing method validity and performance through simulations and real-world data analysis.

Main Results:

• The proposed tests are shown to be valid for detecting novelty.
• The methods successfully identify different types of novelty, including location and scale alternatives.
• Performance evaluation confirms the efficacy of the developed novelty detection techniques.

Conclusions:

• The presented hypothesis tests offer a robust approach to global novelty detection.
• These methods provide a valuable tool for applications where identifying the existence of new patterns is crucial.
• The innovative framework broadens the scope of novelty detection methodologies.