Interpretable transformation and analysis of timelines through learning via surprisability.

Area of Science:

• Data Science
• Computational Statistics
• Cognitive Science

Background:

• High-dimensional timeline data analysis is crucial but challenged by dimensionality, sparsity, and complex distributions.
• Conventional methods struggle to effectively extract insights, identify outliers, and detect anomalies in temporal datasets.
• Human cognitive science highlights 'surprisability' as a key factor in focusing on unexpected deviations.

Purpose of the Study:

• To introduce Learning via Surprisability (LvS), a novel method for transforming high-dimensional timeline data.
• To formalize deviations from expected behavior for quantifying and prioritizing anomalies in time-series data.
• To bridge cognitive attention theories with computational methods for enhanced anomaly detection.

Main Methods:

• Developed the Learning via Surprisability (LvS) approach to transform high-dimensional timeline data.
• Formalized the concept of surprisability to quantify and prioritize anomalies.
• Applied LvS to diverse datasets: sensor data, global mortality statistics, and historical U.S. presidential addresses.

Main Results:

• LvS transformation enables efficient and interpretable identification of outliers and anomalies.
• The method effectively highlights the most variable features within timelines.
• Demonstrated utility across sensor, medical, and historical textual data.

Conclusions:

• Learning via Surprisability (LvS) offers a novel and effective approach for analyzing high-dimensional timeline data.
• LvS provides a new computational lens for interpreting complex datasets by leveraging cognitive principles.
• The method facilitates context-preserving anomaly and outlier detection in diverse temporal data applications.