Porting WEASEL 2.0 to pyts

[aglenis]

Porting WEASEL 2.0 to pyts

[patrickzib]

Thanks for taking the effort!

A couple of comments:

1. n_bins_arg=4
   In WEASEL 2.0 the alphabet size is fixed to 2

See: Alphabet Size

2. window_sizes_arg=[0.1, 0.3, 0.5, 0.7, 0.9]:
   I suppose this means 0.9 times the size of the time series length? Please go for much smaller numbers. I go for, i.e. win_size in np.arange(4, 44) or win_size in np.arange(4, 24). In combination with dilation, this adds up to very large receptive fields.

See: Window sizes

3. first_difference = False
   I do not see the use of first_differences? Please randomly choose from first_differences, too.

See: Ensemble

4. The number of parameter configurations should be between 50 and 100, each choosing from the range of window_sizes, first_differences, and dilation factors.

See: Ensemble

5. WEASEL has a novel feature selection strategy based on variance

See: SFA with Variance

6. strategy_arg='uniform'
   Not sure, what uniform refers to? I am randomly choosing from equi-width and equi-depth

See: [Binning Strategy]
(https://github.com/patrickzib/dictionary/blob/63633eeaa52680f3a1eb016ec95ea0ca2c5430b9/weasel/classification/dictionary_based/_weasel_v2.py#L125)

Hope, this helps. IMO: The most critical parts should be alphabet_size, window-size, differences, and variance in SFA.

[johannfaouzi]

Hi,

Sorry for the delayed response, I saw the notification and forgot about it...

First, thanks @aglenis for the effort and thanks @patrickzib for the feedback! I will need to look at the paper and the source code to provide more detailed, but I will answer some points first.

• Performing dilation just to get the indices sounds suboptimal to me. You can get the indices with a closed formula.
• The default window sizes seem to be from my implementation of WEASEL in pyts (I don't remember the default values in the original implementation of WEASEL, but I prefer in general relative values than absolute values for hyper-parameters).
• The first difference seems to be used with X_train_trend and X_test_trend.
• The strategy argument has different values in pyts: uniform stands for equi-width (the bins all have the same width), while quantile stands for equi-depth (the same number of values fall in each bin).

In general, I like having more hyper-parameters (even if the values are fixed in the original paper) because it might be useful to change these values for other datasets (many people have their own datasets and don't work on the UCR/UEA archive), but I try to keep the default values as close as possible to the ones in the original publication.

I'm very interested in adding WEASEL 2.0 to pyts, so I will further look into your code and also start working on this on my own, and we'll see what we get!

[TonyBagnall]

FYI, WEASEL 2 is in aeon and we have run it to test results
https://github.com/aeon-toolkit/aeon/blob/main/aeon/classification/dictionary_based/_weasel_v2.py