top_kurt_ind = (-kurt).argsort()
# IC
examine_credit_cluster(
X_ica[:, top_kurt_ind[:2]],
y,
title=f"ICA transformation on {data_key} data",
xylabel=["IC1", "IC2"],
fname=f"output/ica_ic_cluster_{data_key}.png",
)
if data_key == "fashion":
# IC
plot_fashion_cluster(
ica.components_[top_kurt_ind[:25], :],
range(25),
fname="output/ica_ic_fashion.png",
)
# reconstructed image
X_recon = ica.inverse_transform(X_ica)
plot_fashion_cluster(X_recon,
y,
fname="output/ica_reconstructed_fashion.png")
# recon vs k
sample = []
for nc in range(2, 201, 8):
print(f"Reconstructing with {nc} PC...")
ica = FastICA(n_components=nc, whiten=True, random_state=0)
X_recon = ica.inverse_transform(ica.fit_transform(X))
plt.savefig(f"output/pca_eigenval_{data_key}.png")
plt.close()
# PC
examine_credit_cluster(
X_pca[:, :2],
y,
title=f"PCA transformation on {data_key} data",
xylabel=["PC1", "PC2"],
fname=f"output/pca_pc_cluster_{data_key}.png",
)
if data_key == "fashion":
# PC image
plot_fashion_cluster(
pca.components_[:25, :], range(25), fname="output/pca_pc_fashion.png"
)
# reconstructed image
pca = PCA(n_components=0.95, whiten=True, random_state=0)
X_recon = pca.inverse_transform(pca.fit_transform(X))
print(f"...Keep {pca.n_components_} components for {data_key} data...")
plot_fashion_cluster(X_recon, y, fname="output/pca_reconstructed_fashion.png")
# recon vs k
sample = []
for nc in range(2, 201, 8):
print(f"Reconstructing with {nc} PC...")
pca = PCA(n_components=nc, whiten=True, random_state=0)
X_recon = pca.inverse_transform(pca.fit_transform(X))
sample.append(X_recon[0])
k_labels_df.to_csv(f"data/kmeans_labels_{data_key}.csv", index=False)
res_df.plot(
subplots=True,
style=".-",
title=f"KMeans performance vs n_clusters on {data_key} data",
)
plt.xlabel("n_clusters")
plt.savefig(f"output/kmeans_{data_key}.png")
plt.close()
if "credit" in data_key:
examine_credit_cluster(
X.values[:, TOP_FEATURES[:2]],
y,
title="True Label",
xylabel=DATA["credit"][0].columns[TOP_FEATURES[:2]],
fname=f"output/true_cluster_{data_key}.png",
)
examine_credit_cluster(
X.values[:, TOP_FEATURES[:2]],
k_labels_df[y.nunique()],
title="KMeans",
xylabel=DATA["credit"][0].columns[TOP_FEATURES[:2]],
fname=f"output/kmeans_cluster_{data_key}.png",
)
if data_key == "fashion":
plot_fashion_cluster(X,
k_labels_df[len(np.unique(y))],
fname="output/kmeans_cluster_fashion.png")
title=f"RCA transformation on {data_key} data",
xylabel=["RC1", "RC2"],
fname=f"output/rca_rc_cluster_{data_key}.png",
)
if data_key == "fashion":
# recon vs k
sample = []
for nc in np.linspace(2, X.shape[1], 25).astype(int):
print(f"Reconstructing with {nc} RC...")
rca = GaussianRandomProjection(n_components=nc, random_state=0)
X_rca = rca.fit_transform(X)
X_recon = np.dot(X_rca, np.linalg.pinv(rca.components_.T))
sample.append(X_recon[0])
plot_fashion_cluster(np.array(sample),
range(25),
fname="output/rca_recon_vs_k_fashion.png")
elif data_key == "credit":
corr_mean = []
corr_std = []
for nc in range(1, X.shape[1]):
print(f"Reconstructing with {nc} RC...")
corr = []
for i in range(100):
rca = GaussianRandomProjection(n_components=nc)
X_rca = rca.fit_transform(X)
X_recon = np.dot(X_rca, np.linalg.pinv(rca.components_.T))
corr.append(
np.corrcoef(X.values.flatten(), X_recon.flatten())[0, 1])
corr_mean.append(np.mean(corr))
corr_std.append(np.std(corr))