MLTeam: Apply LIME for t-SNE.

The note for the meeting of our team is here

(TODO update this figure)

• 09/11/2019: Initial version for simple sampling with noise
• 12/11/2019: Add other sampling methods for image data: perturb image with noise (remove pixel), remove blob(s) of random size.
• 13/11/2019: Add a very simple (naive) explainer using Lasso/ElasticNet
• 14/11/2019: Update plots for some image datasets to show: the scatter plot with samples, the samples input, the "important zone" found by the explainer. (not good for now)

We can add some note for environment configuration, hyperparameters, tip and tricks here.

Ref: https://gist.github.com/gitaarik/8735255

How to: I want to add a repo containing all my simple, common code for any python project. This code is at the py-common repo.

# create a folder in current repo to hold the code in `py-common`
mkdir common

# make the `py-common` repo a submodule in this current repo
git submodule add https://github.com/vu-minh/py-common.git common

# keep the code in `common` up-to-date
git submodule update

Note on the stability of sampling method:

• Update 12/11/2019: the sampling is now NOT STABLE (re-run tsne for a little longer for stable embedding)

Goal: given a selected index of a point selected_idx, sample some points arround it in HD (x_samples) and ask tsne to calculate the embedding of these sampled points in LD (y_samples)

# use the sample function in `sample_tsne` module
from sample_tsne import tsne_sample_embedded_points

# prepare input data `X` and tsne hyper-params
...

# select a point to sample in HD
selected_idx = np.random.randint(X.shape[0])

# create new samples in HD and embed them in LD
Y, x_samples, y_samples = tsne_sample_embedded_points(
    X,
    selected_idx=selected_idx,
    n_samples=n_samples,
    sigma_HD=sigma_HD,
    sigma_LD=sigma_LD,
    tsne_hyper_params=tsne_hyper_params,
    early_stop_hyper_params=early_stop_hyper_params,
)