Hokusai.jl

A julia package for clustering spatial timeseries, like for example eye-tracking data, based on PCCA+.

References

• Bachelor thesis (https://github.com/axsk/bachelor/blob/master/ba.pdf)
• Spectral Clustering for Non-Reversible Markov Chains (https://doi.org/https://doi.org/10.1007/s40314-018-0697-0)

Installation

Requires Julia 0.6.

• Install the package:

Pkg.clone("https://github.com/axsk/hokusai")

• Copy the data file "sallsac_Hokusai.seq" into the packages data directory (~/.julia/v0.6/Hokusai/data).
• Test the package:

Pkg.test("Hokusai")

Basic usage

See the test file

• data is a DataFrame containing the x-/y-coordinates in the first two columns, the relative time in the third and an identifier for aggregation in the fourth column, allowing clustering of multiple test executions.
• n denotes the desired number of clusters.
• tau denotes the timestep used for creating the markov jump process
• sigma denotes the gaussian mixture standard deviation used for the "spatial coupling".
• precluster specifies the number of preclusters to generate via kmeans, to improve performance. (use 0 for no preclustering)
• sort denotes how the final clusters will be ordered. :size sorts by number of fixations in each cluster, and :x by the average horizontal position.
• method specifies the objective function used in the PCCA+ optimization. Accepted values: :scaling (Weber), :metastability (Deuflhard), :crispness (Röblitz)

cluster returns a HokusaiResult, containing the resulting clustering in the field assignments.

todo

• think about pi
• parameter optimization -- automatic tau/gamma -- automatic n (which method in nonreversible case?)

how to choose tau/sigma:

• we want tau to measure every fixation, -> min fixdur?
• sigma should seperate different fixations -> mean fixdist?