summary of GSOC calls

[tdhock]

• template vs subclass for generic loss functions
• generation of docs via roxygen2
• TODOS for next week: investigate if current approach is a good idea with binary segmentation + change in mean loss.
• set.seed(1) to control random numbers (if you use random numbers for input data)

[tdhock]

@diego-urgell during the next call please try to take notes and write them as a new comment in this issue.
yesterday we discussed the new registration in C++ code, and next steps implementing binseg for normal change in mean.

[tdhock]

• implement binary segmentation algorithm, change in mean and constant variance
• more challenging than anticipated. run-time off by one errors.
• summary statistics parent class for cumsum and cumsumsquared
• figure out a way to get summary statistics as a static member of distributions? it can be determined at compile time (does not change at run time)
• compilation warnings remove compilation warnings #4
• add 1 to indices in RcppInterface.cpp before returning to R.
• error handling, number of changepoints too large? invalid distribution or algorithm?
• use tests/testthat/test-*.R files if you want to use devtools::test()
• goals for next week, more test cases for bin seg change in mean. loss and segment mean values. different cost functions. multiple parameter return, change NumericVector before_mean,after_mean to NumericMatrix before_params,after_params
• reading: good intro to ML / stats https://web.stanford.edu/~hastie/ElemStatLearn/

[diego-urgell]

Review:

• Implemented cost function for Normal change in variance and Normal change in mean and variance
• Added minSegLen constraint and a mechanism to avoid zero variance segmentation
• Implemented a mechanism to dynamically create a numeric matrix that contains the necessary parameters, by creating virtual methods and overloading them on each distribution
• Tested the current functions using changepoint

Next Week:

• Create more tests with the three existing distributions (Normal). Use the changepoint package for now.
• Decide and implement a way to determine the cost of each segment from the R interface.
• Set up codecov along with covr for code coverage analysis.
• Implement more distributions and start testing them.
• GSOC mid-term evaluations.

[diego-urgell]

TODOS for next week:

• Correct implementation of Poisson cost function by using MLE for the rate parameter.
• Implement the cost function for the Exponential distribution
• Ask @tdhock about the differences in Negative Binomial cost function and rate parameter estimation between BinSeg and gfpop
• Implement the structure of the S4 class for the user interface
• Set up the coef method to build segmentation models of a particular number of change points

[diego-urgell]

TODOS for this week:

• Implement coef method
• Implement graph method
• Implement resid method
• Finish the user-facing methods and the BinSeg class
• Correct the poisson and exponential cost functions (reduce computing cost)
• Include the missing parts of the log likelihood equation in R code. Achieve -2*logLikfor every distribution.
• Start writing the user documentation.

[rkillick]

Looks like a challenge for this week. Just a quick note that there is no "graph" generic, it is plot.

[diego-urgell]

@rkillick It is definitely a challenge! I worked a lot on the R interface and it is almost complete, in PR #9

After this I will write documentation for the functions :)

[diego-urgell]

TODOS for this week

• Finish the documentation
• Correct the poisson and exponential cost functions (reduce computing cost)
• Include the missing parts of the log likelihood equation in R code. Achieve -2*logLikfor every distribution.
• Make sure the package passes all CRAN checks
• Place static expected result on tests
• Look at Wild Binary Segmentation paper. Start thinking/planning about the implementation

[diego-urgell]

TODOS for this week:

• Correct the poisson and exponential cost functions (reduce computing cost)
• Include the missing parts of the log likelihood equation in R code. Achieve -2*logLik for every distribution except negbin.
• Add tests for the R interface methods
• Check the resid function
• Write a Readme for the Github repository
• Set contributions in Description File
• Setup release for final submission

[diego-urgell]

Hi @rkillick and @tdhock! I finished everything for the final submission! I created a README that has information about the package, and a report on GSOC 2021 at the end. Does it look good?

[tdhock]

yes good job