Sktime: Enhance Proximity Forest

[moradisten]

I've been developing a Proximity Forest Technique in Python based on the authors' project for my degree thesis when I found your toolkit by chance. The thing is, I was testing your Proximity Forest benchmark and I found it's not scalable. It takes too much time for simple datasets. I made a comparison between your Sktime PForest and mine, and I've found mine is much more efficient.

You can check my repository if you want to have a look and don't hesitate if you want me to contribute or give you a hand: https://github.com/moradisten/PForests.git

accuracies.pdf
time_exec.pdf

I'm also developing the LbEnhanced distance measure in Python, based on the project implemented by the original authors in Java: https://github.com/moradisten/KNN-LB.git

Best Regards :)

[TonyBagnall]

hi, thanks for that, I'm tagging George in this who implements PF @goastler , he will take a look. Are you at Monash?

[moradisten]

No :P. I'm currently finishing my degree thesis based on time series
classifiers comparison. I'm a computer science student in the University of
Murcia, Spain

Whatever you need, just say it :)

El vie., 4 sept. 2020 a las 12:16, Tony Bagnall ([e-mail])
escribió:

hi, thanks for that, I'm tagging George in this who implements PF
@goastler https://github.com/goastler , he will take a look. Are you at
Monash?

—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
https://github.com/alan-turing-institute/sktime/issues/388#issuecomment-687058658,
or unsubscribe
https://github.com/notifications/unsubscribe-auth/AHEHQFD5J5LXFQ5RVYJN6WLSEC5ANANCNFSM4QX3PFBA
.

[TonyBagnall]

ah great, I'll let Chang Wei know, he will be pleased :) We would welcome your contributing to sktime, I'll talk to George and Chang Wei and get back to you. We have a wrapped version of his original code and a ported version in tsml, the java toolkit, so we can do a good comparison.

[moradisten]

Alright :)

El vie., 4 sept. 2020 12:28, Tony Bagnall [e-mail]
escribió:

ah great, I'll let Chang Wei know, he will be pleased :) We would welcome
your contributing to sktime, I'll talk to George and Chang Wei and get back
to you. We have a wrapped version of his original code and a ported version
in tsml, the java toolkit, so we can do a good comparison.

—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
https://github.com/alan-turing-institute/sktime/issues/388#issuecomment-687063706,
or unsubscribe
https://github.com/notifications/unsubscribe-auth/AHEHQFA2VJJ4BE6ZXSXUMNDSEC6NLANCNFSM4QX3PFBA
.

[moradisten]

By the way, pleased to meet you. I've read a lot of your articles, which
are pretty interesting and I've referenced some in my report

El vie., 4 sept. 2020 12:29, Morad Abou [e-mail] escribió:

Alright :)

El vie., 4 sept. 2020 12:28, Tony Bagnall [e-mail]
escribió:

ah great, I'll let Chang Wei know, he will be pleased :) We would welcome
your contributing to sktime, I'll talk to George and Chang Wei and get back
to you. We have a wrapped version of his original code and a ported version
in tsml, the java toolkit, so we can do a good comparison.

—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
https://github.com/alan-turing-institute/sktime/issues/388#issuecomment-687063706,
or unsubscribe
https://github.com/notifications/unsubscribe-auth/AHEHQFA2VJJ4BE6ZXSXUMNDSEC6NLANCNFSM4QX3PFBA
.

[fkiraly]

@moradisten, nice to meet you - while you're here and given that you're already working on it for your thesis - would you be interested to implement an overarching distance/kernel interface for sktime, and composition patterns that involve distances/kernels?
I imagine an interface similar to the GP/kernels interface in scikit-learn, and an updated version of #52 (which is a "dormant" piece of work currently).

[moradisten]

Interesting. Tell me more about it

El vie., 4 sept. 2020 a las 14:10, fkiraly ([e-mail])
escribió:

@moradisten https://github.com/moradisten, nice to meet you - while
you're here and given that you're already working on it for your thesis -
would you be interested to implement an overarching distance/kernel
interface for sktime, and composition patterns that involve
distances/kernels?
I imagine an interface similar to the GP/kernels interface in scikit-learn,
and an updated version of #52
https://github.com/alan-turing-institute/sktime/issues/52 (which is a
"dormant" piece of work currently).

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
https://github.com/alan-turing-institute/sktime/issues/388#issuecomment-687104155,
or unsubscribe
https://github.com/notifications/unsubscribe-auth/AHEHQFG5WPYRC4VBIU7HB6LSEDKMFANCNFSM4QX3PFBA
.

[mloning]

@moradisten upgrading the distance based classifiers would be great! They currently don't pass the all of our standard tests and there are some open bug reports on them (see e.g. #413), so that would be a great contribution!

[moradisten]

Perfect, actually I was using that specific classifier (KNN + DTW) for my
thesis, but I didn't know sktime was also implementing it :P
First things first, Id like to know what my first tasks would be because
I'd like to adapt the P. Forest I implemented and check recompare it to the
P.Forest sktime implemented but I'd be delighted to work in the
distance-based classifiers field.

We keep in touch, kind regards :)

El jue., 8 oct. 2020 a las 1:02, Markus Löning ([e-mail])
escribió:

@moradisten https://github.com/moradisten upgrading the distance based
classifiers would be great! They currently don't pass the all of our
standard tests and there are some open bug reports on them (see e.g. #413
https://github.com/alan-turing-institute/sktime/issues/413), so that
would be a great contribution!

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
https://github.com/alan-turing-institute/sktime/issues/388#issuecomment-705237642,
or unsubscribe
https://github.com/notifications/unsubscribe-auth/AHEHQFE3LW62K4ADN4C7YALSJTXSDANCNFSM4QX3PFBA
.

[mloning]

Yes testing your proximity forest against the sktime implementation and potentially also the Java version in TSML would be a good start. I'm tagging @goastler who implemented the sktime and Java version I believe. For KNN, I'm tagging @jasonlines who implemented our current version.

[moradisten]

Great. But I think it would be also necessary to adapt my project to the
sktime syntax. By the way, the Java implementation of PForest is even
faster, which seemed very "funny" even though I did my best to replicate
the same code using python. I think a thourough research on that point
would be interesting. By the way, for DTW I used a very interesting and
very powerful library (
https://dtaidistance.readthedocs.io/en/latest/usage/installation.html) and
I'd be interesting with the DTW your team implemented

As for KNN + DTW, I also implemented the Lb_Enhanced distance researchers
say it's faster than Lb_Keogh, If you guys want to have a look.
(The article: Elastic bands across the path: A new framework and method to
lower bound DTW
https://www.researchgate.net/publication/332892906_Elastic_bands_across_the_path_A_new_framework_and_method_to_lower_bound_DTW
)
(The researchers project in github :
https://github.com/ChangWeiTan/LbEnhanced.git)
(My code: https://github.com/moradisten/KNN-LB.git)

So, we should discuss and organise this matters

Kind Regards

El jue., 8 oct. 2020 12:35, Markus Löning [e-mail]
escribió:

Yes testing your proximity forest against the sktime implementation and
potentially also the Java version in TSML
https://github.com/uea-machine-learning/tsml/tree/master/src/main/java/tsml/classifiers/distance_based/proximity
would be a good start. I'm tagging @goastler https://github.com/goastler
who implemented the sktime and Java version I believe. For KNN, I'm tagging
@jasonlines https://github.com/jasonlines who implemented our current
version.

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
https://github.com/alan-turing-institute/sktime/issues/388#issuecomment-705483422,
or unsubscribe
https://github.com/notifications/unsubscribe-auth/AHEHQFHMBIRYLWIJS7TPNS3SJWIVJANCNFSM4QX3PFBA
.

[wangyida]

@moradisten Aha I also tried your implementation before I found that sktime has already included it. It does take a long time for sktime's API for the benchmark in your repo. I'm also checking to implement TS-CHIEF in such frameworks

[moradisten]

I've been trying, when I had free time, to adapt my project to sktime, but I found the main problem that makes sktime-proximityForest so inefficient is the use of Pandas library. Pandas is great when it comes to dataset management, as far as I've seen, it doesn't perform well when managing time series datasets in Proximity Forest

[mloning]

@wangyida @moradisten We now also support 3d numpy arrays with shape (n_instances, n_variables, n_timepoints), let me know if that helps!

What input format do you have in your implementation?