RobotPsychologist commented 3 weeks ago

Background Research

Completion Deadline: November 20th, 2024

The sktime sibling issue to this one is https://github.com/sktime/sktime/issues/6481, we can add our desired annotation algorithms to that issue as we go.

The lists below are the most heavily cited recent survey papers that may be relevant to our project. After the meta-review, identify newer techniques that reference these survey papers.

Final Deliverable

A prioritized list documented in a markdown file in 0_meal_identification/meal_identification/references

You should record the Abstract Typing, and Metadata for each paper so that the sktime-dev team can properly tag the algorithm in the registry. If you're pressed for time the most important information the, name of the algorithm, literature references, and Abstract Typing.

Sub Deliverables: Create a markdown file for each category of papers below for each paper in a markdown file:

create a brief summary of your thoughts on the paper,
what is useful to us, or is it not useful at all
which algorithms are worth trying, a link to the original paper of the identified algorithm we may want to implement.

Metadata

name of algorithm
literature references
what are hyper-parameters
- clear from paper yes/no
- if yes: name and type if possible
- this can be done after prioritization as this is often complicated (obscure/unclear, requires an understanding of math and algos on formal level)

Abstract Typing

type of label/detection: points, segments, both, something else
- segment = (start time stamp, end time stamp)
- if segments, can they overlap
- are labels deterministic, or probabilistic (probability of segment)
type of label meaning: outlier/anomaly, changepoint, mixed/multiple, sth else
does the algorithm also return a score, e.g., anomaly score
label annotation present? If yes: categorical, numerical
learning type: supervised, unsupervised, semi-supervised
learning mode: stream, batch, both
univariate only or multivariate capability
time series scitype: single time series, panel/collection, hierarchical, multiple of these, sth else

Metrics will have the same dimensions (except perhaps a few), I'll put metrics in a different issue.

Other Information

Implementation/library:

For the algorithms that already have a well-developed implementation or library:

authors (GitHub names etc)
repo URL if applicable
pypi name(s) if applicable
code status: pypi package, contained in package, loose code no package, only paper (no code)
language: python, R, julia, sth else
license (type): copyleft, permissive, copyrighted
maintenance status: still maintained, no longer maintained, unclear
governance model, who owns/controls, key points of contact (e.g., maintainers, company)

Packages That Contain Detectors

They are both active and defunct that contain detectors (without the typing typically!)

ruptures
adtk
- Arundo toolbox - is that the same?
gluon-ts detection module
seqlearn
seglearn (one of the two has detectors but I forgot which)
darts detection module
pyod, but that is tabular (non-temporal)
hmmlearn
- can do supervised and unsupervised, single series and panel/collection
- but sktime has only one interfaced
- and hmmlearn provides only variants of one major algorithm
Turing change point detection benchmark
- does not have cleanly delineated algorithms, but a larger list of names
- snippets of code many of which are not reusable in multiple languages
numenta benchmark

Papers

Change Point Detection

[ ] #169
[ ] #168
[x] #138

Benchmarks

For these benchmarks, assess which are most similar to our meal detection problem, and then see which algorithms are currently performing best on the most appropriate benchmarks to include in sktime.