think about time series similarity

[aronwc]

Good empirical survey of the (many) time series similarity functions:

• http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.140.8203&rep=rep1&type=pdf
• comparisons on 38 different datasets suggest Dynamic Time Warping is often the best for classification

[aronwc]

Another good survey:

• Clustering of time series data—a survey

[aronwc]

Another time series thought: There's a lot of work analyzing physiology data (e.g., heart rate). One popular version of this is something like:

• Use a sliding window of fixed size to produce a list of subsequences
• Cluster the windows
• Use these clusters as features to predict something (e.g., heart failure in 3 hours)

I could imagine a version of this for predicting flu outbreak or other health things.