jylfamily/interpert_machine_learning

https://christophm.github.io/interpretable-ml-book/index.html

interpert_machine_learning's Issues

"Resolving Discrepancies in Interaction Bins for EBM Model: Approach and Query"

In the context of my machine learning project, I'm currently working on generating a heatmap for the interaction functions within the EBM model. The focus of this model is on predicting baryon fractions, and I'm particularly interested in visualizing the interaction effects based on pairs of parameters. To create this heatmap, I've utilized the pcolormesh technique.

However, I've encountered a discrepancy in the number of interaction bins. The EBM model was configured with a maximum of 32 interaction bins, yet upon inspecting ebm.bins_, I've noticed the presence of only 29 interaction bins. The EBM documentation doesn't provide a clear explanation for the absence of these 3 bins. To address this, I've attempted to align the shape of term_scores_ by performing linear interpolation on the bins. While I've taken this approach, I'm uncertain if it's the correct way to address the issue.

Could you kindly guide me on the appropriate method to handle this situation?
And here is the code for the pcolormesh:

def add_edges(edges, num_to_add):
"""Add additional bin edges through linear interpolation."""
last_diff = edges[-1] - edges[-2]
return np.concatenate([edges, [edges[-1] + last_diff * (i+1) for i in range(num_to_add)]])

it_feat = [f for f in ebm.term_features_ if len(f) > 1]
fig, ax = plt.subplots(nrows=2, ncols=5, figsize=(22, 8))
j = 0
it_scores = ebm.term_scores_[10:]

Define a common color range for all plots
vmin = min([score.min() for score in it_scores])
vmax = max([score.max() for score in it_scores])

for row in ax:
for col in row:
idx, idy = it_feat[j]

x_edges = add_edges(ebm.bins_[idx][1], 4)
y_edges = add_edges(ebm.bins_[idy][1], 4)
print(x_edges.shape)
c = col.pcolormesh(x_edges, y_edges, it_scores[j], shading='auto', cmap='bwr', vmin=vmin, vmax=vmax)
colnameX, colnameY = ebm.feature_names_in_[it_feat[j][0]], ebm.feature_names_in_[it_feat[j][1]]
col.set_xlabel(colnameX)
col.set_ylabel(colnameY)

j += 1

plt.tight_layout(pad=2.5)

Add a common colorbar
fig.subplots_adjust(right=0.83)
cbar_ax = fig.add_axes([0.85, 0.15, 0.01, 0.7])
fig.colorbar(c, cax=cbar_ax, label='Baryon Fraction')

fig.savefig('interaction_term_score.png')
plt.show()