def _title_from_terms(terms, ic_idx, label=None, n_terms=4, sign=1):
if terms is None:
return '%s[%d]' % (label, ic_idx)
# Use the n terms weighted most as a positive title, n terms
# weighted least as a negative title and return both
pos_terms = get_n_terms(terms, ic_idx, n_terms=n_terms, sign=sign)
neg_terms = get_n_terms(terms, ic_idx, n_terms=n_terms, top_bottom="bottom",
sign=sign)
title = '%s[%d]: POS(%s) \n NEG(%s)' % (
label, ic_idx, ', '.join(pos_terms), ', '.join(neg_terms))
return title
def plot_term_comparisons(terms, labels, ic_idx_list, sign_list, color_list=('g', 'r', 'b'),
top_n=4, bottom_n=4, standardize=True, out_dir=None):
"""
Take the list of ica image terms and the indices of components to be compared, and
plots the top n and bottom n term values for each component as a radar graph.
The sign_list should indicate whether term values should be flipped (-1) or not (1).
"""
assert len(terms) == len(labels)
assert len(terms) == len(ic_idx_list)
assert len(terms) == len(sign_list)
assert len(terms) == len(color_list)
n_comp = len(ic_idx_list[0]) # length of each ic_idx_list and sign_list
for i in range(len(terms)):
assert len(ic_idx_list[i]) == n_comp
assert len(sign_list[i]) == n_comp
# iterate over the ic_idx_list and sign_list for each term and plot
# store top n and bottom n terms for each label
term_arr = np.empty((len(labels), n_comp, top_n + bottom_n), dtype="S30")
for n in range(n_comp):
terms_of_interest = []
term_vals = []
name = ''
for i, (term, label) in enumerate(zip(terms, labels)):
idx = ic_idx_list[i][n]
sign = sign_list[i][n]
# Get list of top n and bottom n terms for each term list
top_terms = get_n_terms(
term, idx, n_terms=top_n, top_bottom='top', sign=sign)
bottom_terms = get_n_terms(
term, idx, n_terms=bottom_n, top_bottom='bottom', sign=sign)
combined = np.append(top_terms, bottom_terms)
terms_of_interest.append(combined)
term_arr[i][n] = combined
# Also store term vals (z-score if standardize) for each list
t, vals = get_ic_terms(term, idx, sign=sign, standardize=standardize)
s = pd.Series(vals, index=t, name=label)
term_vals.append(s)
# Construct name for the comparison
name += label + '[%d] ' % (idx)
# Data for all the terms
termscore_df = pd.concat(term_vals, axis=1)
# Get unique terms from terms_of_interest list
toi_unique = np.unique(terms_of_interest)
# Get values for unique terms_of_interest
data = termscore_df.loc[toi_unique]
data = data.sort_values(list(labels), ascending=False)
# Now plot radar!
N = len(toi_unique)
theta = radar_factory(N)
fig = plt.figure(figsize=(10, 8))
ax = fig.add_subplot(1, 1, 1, projection='radar')
title = "Term comparisons for %scomponents" % (name)
ax.set_title(title, weight='bold', size='medium', position=(0.5, 1.1),
horizontalalignment='center', verticalalignment='center')
y_min, y_max, y_tick = nice_bounds(data.values.min(), data.values.max())
ax.set_ylim(y_min, y_max)
ax.set_yticks([0], minor=True)
ax.set_yticks(y_tick)
ax.yaxis.grid(which='major', linestyle=':')
ax.yaxis.grid(which='minor', linestyle='-')
for label, color in zip(labels, color_list):
ax.plot(theta, data[label], color=color)
ax.fill(theta, data[label], facecolor=color, alpha=0.25)
ax.set_varlabels(data.index.values)
legend = plt.legend(labels, loc=(1.1, 0.9), labelspacing=0.1)
plt.setp(legend.get_texts(), fontsize='small')
# plt.show()
# Saving
if out_dir is not None:
save_and_close(
out_path=op.join(out_dir, '%sterm_comparisons.png' % (
name.replace(" ", "_"))))
# Save top n and bottom n terms for each label
term_dfs = []
col_names = ["top%d" % (n + 1) for n in range(top_n)] + ["bottom%d" % (n + 1) for n in range(bottom_n)]
for i, label in enumerate(labels):
term_df = pd.DataFrame(term_arr[i], columns=["%s_%s" % (label, col) for col in col_names])
term_df.insert(0, "%s_idx" % (label), ic_idx_list[i])
term_dfs.append(term_df)
all_term_df = pd.concat(term_dfs, axis=1)
out_file = op.join(out_dir, 'term_comparison_summary.csv')
all_term_df.to_csv(out_file)