AttrLLM / visualization /plotting /interactions.py
Qingpeng Kong
clean initial state
3e72399
"""
Plotting utilities for visualizing higher-order feature interactions in a
text-interpretability web app.
Inputs are assumed to come from attribution utilities such as
`shapley_interactions(...)` or `banzhaf_interactions(...)`.
Outputs are Plotly Figure objects that can be rendered directly in Gradio/UI.
"""
import json
import math
import uuid
from collections import defaultdict
from html import escape
import plotly.graph_objects as go
import numpy as np
from typing import List, Tuple, Dict, Optional
try: # optional dependency
from shapiq.interaction_values import InteractionValues # type: ignore
from shapiq.plot import bar_plot # type: ignore
except Exception: # pragma: no cover
InteractionValues = None
bar_plot = None
from .utils import format_feature_label, get_color_scale, create_legend, matplotlib_to_plotly
_TOKEN_VIEW_STYLE = """
<style>
.token-interaction-view {
--token-border: #e0d9f0;
--token-active: #7048e8;
--token-bg: #faf8ff;
--panel-bg: linear-gradient(135deg, #f6f3ff 0%, #f0f4ff 100%);
font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", "Helvetica Neue", Arial, sans-serif;
background: var(--token-bg);
border: 2px solid var(--token-border);
border-radius: 20px;
padding: 24px;
display: flex;
flex-direction: column;
gap: 20px;
box-shadow: 0 4px 20px rgba(88, 60, 140, 0.08);
margin: 16px 0;
}
.token-interaction-panel {
background: var(--panel-bg);
border-radius: 16px;
padding: 16px 20px;
border: 1.5px solid rgba(112, 72, 232, 0.15);
font-size: 14px;
font-weight: 500;
color: #2d1f4a;
line-height: 1.6;
box-shadow: 0 2px 12px rgba(112, 72, 232, 0.06);
}
.token-interaction-grid {
display: flex;
flex-wrap: wrap;
gap: 16px;
}
.interaction-token {
border-radius: 16px;
border: 2px solid #e7e2f5;
padding: 14px 16px;
width: min(260px, 100%);
background: linear-gradient(135deg, #ffffff 0%, #faf8ff 100%);
transition: all 0.3s cubic-bezier(0.34, 1.56, 0.64, 1);
cursor: pointer;
box-shadow: 0 2px 8px rgba(88, 60, 140, 0.06);
}
.interaction-token:hover {
transform: translateY(-2px);
box-shadow: 0 8px 24px rgba(112, 72, 232, 0.15);
border-color: rgba(112, 72, 232, 0.3);
}
.interaction-token[open] {
border-color: var(--token-active);
box-shadow: 0 12px 32px rgba(112, 72, 232, 0.22);
transform: translateY(-3px);
background: linear-gradient(135deg, #ffffff 0%, #f6f3ff 100%);
}
.interaction-token summary {
list-style: none;
cursor: pointer;
display: flex;
flex-direction: column;
gap: 6px;
}
.interaction-token summary::-webkit-details-marker {
display: none;
}
.interaction-token__score {
font-size: 13px;
font-weight: 700;
letter-spacing: 0.03em;
color: #7048e8;
background: linear-gradient(135deg, #7048e8 0%, #9b6dff 100%);
-webkit-background-clip: text;
-webkit-text-fill-color: transparent;
background-clip: text;
}
.interaction-token__text {
font-size: 14px;
font-weight: 500;
color: #2d1f4a;
white-space: normal;
overflow: hidden;
text-overflow: ellipsis;
line-height: 1.5;
}
.interaction-token__hint {
font-size: 12px;
color: #7a6b99;
font-weight: 500;
margin-top: 2px;
}
.token-link-list {
list-style: none;
padding: 12px 0 0 0;
margin: 12px 0 0 0;
border-top: 2px solid rgba(112, 72, 232, 0.1);
}
.token-link-list li {
display: flex;
justify-content: space-between;
align-items: center;
padding: 8px 0;
font-size: 13px;
}
.token-link-list li + li {
border-top: 1.5px dashed rgba(112, 72, 232, 0.12);
}
.token-link-name {
color: #3a2f50;
font-weight: 500;
margin-right: 12px;
flex: 1;
}
.token-link-value {
font-weight: 700;
font-size: 13px;
color: #7048e8;
background: rgba(112, 72, 232, 0.08);
padding: 4px 10px;
border-radius: 999px;
}
.token-interaction-empty {
font-size: 13px;
font-weight: 500;
color: #9a8bb5;
margin-top: 8px;
font-style: italic;
}
@media (prefers-color-scheme: dark) {
.token-interaction-view {
--token-border: #2d3c53;
--token-active: #c7d2ff;
--token-bg: #0f1522;
--panel-bg: linear-gradient(135deg, #141f33 0%, #19263d 100%);
color: #edf2ff;
box-shadow: 0 12px 32px rgba(0, 0, 0, 0.28);
}
.token-interaction-panel,
.interaction-token,
.sentence-row {
color: #edf2ff;
background: linear-gradient(135deg, #152033 0%, #101929 100%);
border-color: rgba(199, 210, 255, 0.14);
}
.interaction-token[open] {
background: linear-gradient(135deg, #18243a 0%, #121b2c 100%);
border-color: #c7a7ff;
box-shadow: 0 14px 34px rgba(115, 76, 255, 0.24);
}
.interaction-token__text,
.token-link-name,
.sentence-row__text,
.sentence-row__score,
.sentence-interaction-header,
.sentence-link-label {
color: #edf2ff;
}
.interaction-token__hint,
.token-interaction-empty,
.sentence-row__links {
color: #9fb0c9;
}
.token-link-value,
.sentence-row__badge {
background: rgba(160, 178, 255, 0.14);
color: #dbe4ff;
}
.sentence-interaction-view {
--token-border: #2d3c53;
--token-bg: #0f1522;
}
}
.sentence-interaction-view {
--token-border: #d9d5e0;
--token-bg: #fbf8ff;
display: flex;
flex-direction: column;
gap: 16px;
padding: 18px;
border: 1px solid var(--token-border);
border-radius: 18px;
background: var(--token-bg);
font-family: "Segoe UI", "Helvetica Neue", Arial, sans-serif;
}
.sentence-interaction-header {
font-size: 14px;
font-weight: 600;
color: #4a3c71;
}
.sentence-interaction-body {
display: grid;
grid-template-columns: minmax(0, 1fr) 240px;
gap: 24px;
}
.sentence-list {
display: flex;
flex-direction: column;
gap: 10px;
}
.sentence-row {
display: grid;
grid-template-columns: 90px minmax(0, 1fr);
gap: 16px;
align-items: center;
padding: 10px 14px;
border-radius: 14px;
border: 1px solid var(--token-border);
background: #fff;
box-shadow: 0 8px 18px rgba(64, 58, 95, 0.08);
height: 64px;
}
.sentence-row__score {
font-size: 13px;
font-weight: 600;
color: #2a1f44;
}
.sentence-row__text {
font-size: 13px;
color: #2c233b;
white-space: nowrap;
overflow: hidden;
text-overflow: ellipsis;
}
.sentence-row__links {
grid-column: 1 / -1;
font-size: 11px;
color: #6d6483;
display: flex;
gap: 6px;
flex-wrap: wrap;
}
.sentence-row__badge {
background: rgba(112, 72, 232, 0.08);
border-radius: 999px;
padding: 2px 8px;
font-weight: 600;
color: #5f43c2;
}
.sentence-links {
position: relative;
min-height: 160px;
}
.sentence-links svg {
width: 100%;
height: var(--canvas-height, 200px);
overflow: visible;
}
.sentence-link-path {
fill: none;
opacity: 0.9;
}
.sentence-link-label {
font-size: 11px;
fill: #4b3f66;
}
.sentence-node {
fill: #fff;
stroke: #c3bed7;
stroke-width: 1.5;
}
</style>
"""
def _strip_occurrence_suffix(text: str) -> str:
text = text or ""
if text.endswith(")") and " (" in text:
base, _, tail = text.rpartition(" (")
if tail[:-1].isdigit():
return base
return text
_NETWORK_VIEW_STYLE = """
<style>
.interaction-network {
position: relative;
background: linear-gradient(135deg, #faf8ff 0%, #f0f4ff 100%);
border: 2px solid #e0d9f0;
border-radius: 20px;
padding: 20px 24px 16px;
box-shadow: 0 8px 32px rgba(88, 60, 140, 0.12), 0 2px 8px rgba(88, 60, 140, 0.08);
font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", "Helvetica Neue", Arial, sans-serif;
margin: 16px 0;
}
.network-toolbar {
display: flex;
justify-content: space-between;
align-items: center;
gap: 16px;
margin-bottom: 12px;
color: #2d1f4a;
}
.network-title {
font-weight: 700;
font-size: 18px;
letter-spacing: -0.01em;
color: #2d1f4a;
background: linear-gradient(135deg, #7048e8 0%, #9b6dff 100%);
-webkit-background-clip: text;
-webkit-text-fill-color: transparent;
background-clip: text;
}
.network-hint {
font-size: 13px;
font-weight: 500;
color: #5b4a7a;
background: linear-gradient(135deg, rgba(112, 72, 232, 0.12) 0%, rgba(155, 109, 255, 0.12) 100%);
padding: 8px 16px;
border-radius: 999px;
border: 1.5px solid rgba(112, 72, 232, 0.18);
box-shadow: 0 2px 8px rgba(112, 72, 232, 0.08);
}
.network-svg {
width: 100%;
height: 580px;
border-radius: 16px;
background: #ffffff;
border: 2px solid #ebe7f5;
cursor: pointer;
box-shadow: inset 0 2px 8px rgba(112, 72, 232, 0.04);
}
@keyframes fadeIn {
from { opacity: 0; transform: scale(0.96) translateY(4px); }
to { opacity: 1; transform: scale(1) translateY(0); }
}
@keyframes pulse {
0%, 100% { transform: scale(1); }
50% { transform: scale(1.05); }
}
.network-node {
animation: fadeIn 0.4s cubic-bezier(0.34, 1.56, 0.64, 1);
cursor: pointer;
transition: all 0.3s cubic-bezier(0.34, 1.56, 0.64, 1);
}
.network-node:hover {
animation: pulse 0.6s ease-in-out infinite;
}
.network-edge {
animation: fadeIn 0.5s ease-out;
stroke-linecap: round;
opacity: 0.88;
transition: all 0.3s ease;
}
.edge-label {
font-size: 12px;
font-weight: 700;
fill: #3a2858;
paint-order: stroke;
stroke: #ffffff;
stroke-width: 4px;
text-anchor: middle;
opacity: 0.95;
}
.network-node text {
font-size: 13px;
fill: #1a0f2e;
pointer-events: none;
font-weight: 600;
}
.network-node circle {
stroke-width: 2.5px;
filter: drop-shadow(0 4px 16px rgba(112, 72, 232, 0.2));
transition: all 0.3s cubic-bezier(0.34, 1.56, 0.64, 1);
}
.network-node:hover circle {
filter: drop-shadow(0 8px 24px rgba(112, 72, 232, 0.35));
stroke-width: 3px;
}
.network-status {
font-size: 13px;
font-weight: 500;
color: #5b4a7a;
margin-top: 12px;
padding: 10px 16px;
background: linear-gradient(135deg, rgba(112, 72, 232, 0.06) 0%, rgba(155, 109, 255, 0.06) 100%);
border-radius: 12px;
border: 1.5px solid rgba(112, 72, 232, 0.12);
text-align: center;
box-shadow: 0 2px 8px rgba(112, 72, 232, 0.06);
}
.network-empty {
padding: 20px 16px;
border-radius: 14px;
background: linear-gradient(135deg, #fff5f5 0%, #ffe8e8 100%);
border: 2px solid #ffd0d0;
color: #9a2a42;
font-size: 14px;
font-weight: 500;
text-align: center;
}
.interaction-fallback {
position: relative;
}
</style>
"""
def _interactions_to_shapiq(
interactions: List[Tuple[Tuple[str, ...], float]],
method: str,
order: int,
) -> Tuple[Optional["InteractionValues"], List[str]]:
if InteractionValues is None or not interactions:
return None, []
feature_index: Dict[str, int] = {}
feature_names: List[str] = []
def _idx(name: str) -> int:
if name not in feature_index:
feature_index[name] = len(feature_names)
feature_names.append(name)
return feature_index[name]
lookup: Dict[Tuple[int, ...], int] = {}
values: List[float] = []
for feats, value in interactions:
if not feats:
continue
idx_tuple = tuple(sorted(_idx(f) for f in feats))
lookup[idx_tuple] = len(values)
values.append(float(value))
if not values:
return None, feature_names
if order == 1:
index = "SV" if method == "shapley" else ("IV" if method == "influence" else "BV")
else:
index = "SII" if method == "shapley" else ("III" if method == "influence" else "BII")
min_order = 1 if order <= 1 else order
iv = InteractionValues(
values=np.array(values, dtype=float),
index=index,
max_order=order,
n_players=len(feature_names),
min_order=min_order,
interaction_lookup=lookup,
estimated=False,
baseline_value=0.0,
)
return iv, feature_names
def plot_top_interactions(
interactions: List[Tuple[Tuple[str, ...], float]],
top_k: int = 10,
order: int = 2,
method: str = "shapley"
) -> go.Figure:
"""
Visualize the top-k interactions as a bar chart.
Args:
interactions: List of interactions from shapley_interactions/banzhaf_interactions.
Each item is ((feature_name, ...), value).
top_k: Number of top interactions to display.
order: Interaction order (2 or 3).
method: Attribution method label ("shapley" or "banzhaf").
Returns:
Plotly Figure.
Example:
# From attribution
mobius = run_proxyspex(set_function, features, max_order=3)
interactions = shapley_interactions(mobius, order=2, top_k=10)
fig = plot_top_interactions(interactions, top_k=10, order=2, method="shapley")
"""
if not interactions:
return go.Figure().update_layout(
title="No interactions available",
template="plotly_white"
)
ranked = sorted(interactions, key=lambda item: abs(item[1]), reverse=True)[:top_k]
if bar_plot is not None:
iv, feature_names = _interactions_to_shapiq(ranked, method, order)
if iv is not None:
ax = bar_plot(
[iv],
feature_names=feature_names,
show=False,
abbreviate=False,
max_display=top_k,
global_plot=True,
plot_base_value=True,
)
fig = ax.figure if ax is not None else None
if fig is not None:
return matplotlib_to_plotly(
fig,
title=f"Top {len(ranked)} order-{order} {method.title()} interactions",
)
labels = [
format_feature_label(" · ".join(feats), max_length=50)
for feats, _ in ranked
]
values = [val for _, val in ranked]
# Influence scores are always non-negative (squared Fourier coefficients)
is_influence = method.lower() == "influence"
if is_influence:
values = [abs(v) for v in values]
# Create color scale based on value magnitude (importance)
max_abs_val = max(abs(v) for v in values) if values else 1.0
def get_color(val: float) -> str:
"""Map value to color: purple = positive, red = negative (matches legend)."""
norm = abs(val) / max_abs_val if max_abs_val > 0 else 0.5
if val >= 0:
# Positive: Lavender -> deep violet
r = int(76 + (214 - 76) * (1 - norm))
g = int(29 + (190 - 29) * (1 - norm))
b = int(149 + (255 - 149) * (1 - norm))
else:
# Negative: Light rose -> deep red
r = int(139 + (255 - 139) * (1 - norm))
g = int(0 + (160 - 0) * (1 - norm))
b = int(0 + (122 - 0) * (1 - norm))
return f"rgb({r}, {g}, {b})"
colors = [get_color(v) for v in values]
fig = go.Figure(
data=[
go.Bar(
y=list(reversed(labels)),
x=list(reversed(values)),
orientation="h",
marker=dict(
color=list(reversed(colors)),
line=dict(color="#f5f5f5", width=2),
),
text=[f"{v:.3f}" for v in reversed(values)],
textposition="outside",
textfont=dict(size=14, weight='bold'),
cliponaxis=False,
hovertemplate="<b>%{y}</b><br>Value: %{x:.4f}<extra></extra>",
)
]
)
if not is_influence:
fig.add_vline(x=0, line_dash="dash", line_color="#8c8c8c", line_width=2)
x_axis_label = "Influence Magnitude" if is_influence else "Contribution"
annotation_text = (
"Influence scores are always non-negative (squared Fourier coefficients). Color intensity shows magnitude."
if is_influence else
"Color intensity shows interaction strength: red = negative, purple = positive."
)
fig.update_layout(
title=dict(
text=f"Top {len(labels)} order-{order} {method.title()} interactions",
font=dict(size=18, weight='bold')
),
xaxis_title=dict(text=x_axis_label, font=dict(size=14)),
yaxis_title=None,
xaxis=dict(
tickfont=dict(size=12),
gridcolor="rgba(148, 163, 184, 0.18)",
zerolinecolor="rgba(148, 163, 184, 0.28)",
rangemode="tozero" if is_influence else "normal",
),
yaxis=dict(tickfont=dict(size=13), automargin=True),
template="none",
paper_bgcolor="rgba(0,0,0,0)",
plot_bgcolor="rgba(0,0,0,0)",
font=dict(color="#2d1f4a"),
hovermode="y",
legend=create_legend(method, order),
margin=dict(l=140, r=20, t=70, b=120),
height=max(800, 70 * len(labels)), # Much larger: 800px minimum, 70px per bar
annotations=[
dict(
text=annotation_text,
xref="paper",
yref="paper",
x=0.5,
y=-0.08,
showarrow=False,
font=dict(size=11, color="#666"),
xanchor='center',
yanchor='top',
)
],
)
return fig
def plot_interaction_network(
features: List[str],
interactions: List[Tuple[Tuple[str, ...], float]],
threshold: float = 0.1
) -> go.Figure:
"""
Visualize pairwise interactions as a network graph.
Args:
features: The full ordered feature list.
interactions: Pairwise interactions [((feat_i, feat_j), value), ...].
threshold: Only show interactions with absolute value greater than this threshold.
Returns:
Plotly Figure (network-style visualization).
"""
if not interactions:
return go.Figure().update_layout(
title="No pairwise interactions available",
template="plotly_white"
)
filtered = [
item for item in interactions if abs(item[1]) >= threshold
]
if not filtered:
return go.Figure().update_layout(
title=f"No interactions exceed threshold ({threshold})",
template="plotly_white"
)
n = len(features)
angles = np.linspace(0, 2 * np.pi, max(n, 1), endpoint=False)
positions = {
feat: (float(np.cos(theta)), float(np.sin(theta)))
for feat, theta in zip(features, angles)
}
max_abs = max(abs(val) for _, val in filtered) or 1.0
traces = []
for (feat_a, feat_b), value in filtered:
if feat_a not in positions or feat_b not in positions:
continue
(x0, y0), (x1, y1) = positions[feat_a], positions[feat_b]
color = "#d73027" if value >= 0 else "#4575b4"
width = 1 + 4 * abs(value) / max_abs
label = f"{feat_a} <-> {feat_b}: {value:.3f}"
traces.append(
go.Scatter(
x=[x0, x1],
y=[y0, y1],
mode="lines",
line=dict(color=color, width=width),
hoverinfo="text",
text=[label, label],
showlegend=False,
)
)
node_trace = go.Scatter(
x=[positions[f][0] for f in features if f in positions],
y=[positions[f][1] for f in features if f in positions],
mode="markers+text",
marker=dict(size=14, color="#4a4a4a", line=dict(width=2, color="#ffffff")),
text=[format_feature_label(f, 18) for f in features if f in positions],
textposition="bottom center",
hoverinfo="text",
)
fig = go.Figure(data=traces + [node_trace])
fig.update_layout(
title="Pairwise interaction network",
xaxis=dict(visible=False),
yaxis=dict(visible=False),
template="plotly_white",
showlegend=False,
margin=dict(l=20, r=20, t=60, b=20),
)
return fig
def plot_interaction_matrix(
features: List[str],
interactions: List[Tuple[Tuple[int, int], float]]
) -> go.Figure:
"""
Visualize pairwise interactions as a matrix heatmap.
Args:
features: Ordered feature list (labels for axes).
interactions: List of ((i, j), value) where i/j are feature indices.
Returns:
Plotly Figure (heatmap).
"""
n = len(features)
if n == 0:
return go.Figure().update_layout(
title="Interaction matrix (no features)",
template="plotly_white"
)
matrix = np.zeros((n, n), dtype=float)
for (i, j), value in interactions:
if 0 <= i < n and 0 <= j < n:
matrix[i, j] = value
matrix[j, i] = value
heatmap = go.Heatmap(
z=matrix,
x=[format_feature_label(f, 18) for f in features],
y=[format_feature_label(f, 18) for f in features],
colorscale=get_color_scale("shapley"),
colorbar=dict(title="Interaction value"),
hovertemplate=" %{y} vs %{x}<br>value=%{z:.2e}<extra></extra>",
)
fig = go.Figure(data=[heatmap])
fig.update_layout(
title="Pairwise interaction matrix",
xaxis=dict(side="top"),
yaxis=dict(autorange="reversed"),
template="plotly_white",
margin=dict(l=80, r=20, t=80, b=80),
)
return fig
def plot_3rd_order_interactions(
interactions: List[Tuple[Tuple[str, ...], float]],
top_k: int = 5
) -> go.Figure:
"""
Visualize third-order (triplet) interactions.
Uses grouped or stacked bars to show the top-k triplets.
Args:
interactions: Triplet interactions [((f1, f2, f3), value), ...].
top_k: Number of top triplets to plot.
Returns:
Plotly Figure.
"""
if not interactions:
return go.Figure().update_layout(
title="No third-order interactions available",
template="plotly_white"
)
ranked = sorted(interactions, key=lambda item: abs(item[1]), reverse=True)[:top_k]
labels = [
format_feature_label(" · ".join(feats), max_length=40)
for feats, _ in ranked
]
values = [val for _, val in ranked]
colors = ["#d73027" if v >= 0 else "#4575b4" for v in values]
fig = go.Figure(
data=[
go.Bar(
x=labels,
y=values,
marker=dict(color=colors),
text=[f"{v:.3f}" for v in values],
textposition="auto",
)
]
)
fig.update_layout(
title=f"Top {len(labels)} third-order interactions",
xaxis_title="Feature triplet",
yaxis_title="Interaction value",
template="plotly_white",
margin=dict(l=60, r=20, t=60, b=100),
)
return fig
def _token_colors(value: float, max_abs: float) -> Tuple[str, str]:
if max_abs <= 0:
return "rgba(229, 226, 240, 0.7)", "rgba(193, 189, 209, 0.9)"
norm = max(-1.0, min(1.0, value / max_abs))
if norm >= 0:
base = (222, 86, 61)
else:
base = (47, 128, 237)
norm = -norm
neutral = (245, 242, 252)
r = int(round(neutral[0] + (base[0] - neutral[0]) * norm))
g = int(round(neutral[1] + (base[1] - neutral[1]) * norm))
b = int(round(neutral[2] + (base[2] - neutral[2]) * norm))
alpha = 0.35 + 0.4 * norm
return f"rgba({r}, {g}, {b}, {alpha:.3f})", f"rgb({r}, {g}, {b})"
def _wrap(text: str, max_len: int = 20) -> List[str]:
"""Wrap text into lines of max_len characters"""
if not text or len(text) <= max_len:
return [text]
words = text.split()
lines = []
current_line = []
current_length = 0
for word in words:
word_len = len(word)
if current_length + word_len + len(current_line) > max_len:
if current_line:
lines.append(' '.join(current_line))
current_line = [word]
current_length = word_len
else:
# Single word longer than max_len
lines.append(word[:max_len])
current_line = []
current_length = 0
else:
current_line.append(word)
current_length += word_len
if current_line:
lines.append(' '.join(current_line))
return lines if lines else [text[:max_len]]
def create_interaction_token_view(
features: List[str],
feature_values: List[float],
pairwise: List[Tuple[Tuple[str, ...], float]],
method: str = "shapley",
max_links: int = 5,
layout: str = "token",
) -> str:
"""
Render token interactions as a lightweight chip list.
"""
if not features:
return "<div class='token-interaction-empty'>No tokens available.</div>"
values = list(feature_values) if feature_values else []
if len(values) < len(features):
values.extend([0.0] * (len(features) - len(values)))
# partner lookup for the always-visible chip list
adjacency: Dict[str, List[Tuple[str, float]]] = defaultdict(list)
for feats, val in pairwise:
if len(feats) != 2:
continue
a, b = feats
adjacency[a].append((b, float(val)))
adjacency[b].append((a, float(val)))
for key in adjacency:
adjacency[key].sort(key=lambda item: abs(item[1]), reverse=True)
feature_index = {feat: idx for idx, feat in enumerate(features)}
edges: List[Tuple[int, int, float]] = []
for feats, val in pairwise:
if len(feats) != 2:
continue
a_idx = feature_index.get(feats[0])
b_idx = feature_index.get(feats[1])
if a_idx is None or b_idx is None or a_idx == b_idx:
continue
edges.append((a_idx, b_idx, float(val)))
# Fallback: if no edges are provided, synthesize simple neighbor links so the UI isn't empty.
if not edges and len(features) > 1:
for i in range(len(features) - 1):
score = 0.5 * (values[i] + values[i + 1])
edges.append((i, i + 1, float(score)))
max_abs = max((abs(v) for v in values), default=0.0) or 1.0
chips_html = _render_token_chip_view(
features,
values,
adjacency,
method,
max_abs,
max_links,
)
return chips_html
def _render_interaction_network(
features: List[str],
values: List[float],
edges: List[Tuple[int, int, float]],
method: str,
) -> go.Figure:
"""
Interactive Plotly network view showing pairwise feature interactions.
Args:
features: List of token/feature labels
values: Attribution values for each feature
edges: List of (source_idx, target_idx, interaction_weight) tuples
method: Attribution method name (for display)
Returns:
Plotly Figure with interactive network graph
"""
if not features or not edges:
# Return empty figure with message
fig = go.Figure()
fig.update_layout(
title=f"{method.title()} pairwise interactions",
annotations=[{
"text": "No interaction data available",
"xref": "paper",
"yref": "paper",
"x": 0.5,
"y": 0.5,
"showarrow": False,
"font": {"size": 16, "color": "#666"}
}],
template="plotly_white",
height=600,
)
return fig
# Prepare node data
max_abs_value = max((abs(v) for v in values), default=0.0) or 1.0
# Sort edges by absolute weight and limit to top 60
edges_sorted = sorted(edges, key=lambda item: abs(item[2]), reverse=True)[:60]
# Calculate circular layout positions
n = len(features)
angle_step = 2 * math.pi / n
radius = 100
node_positions = {}
for idx in range(n):
angle = idx * angle_step - math.pi / 2 # Start from top
node_positions[idx] = {
'x': radius * math.cos(angle),
'y': radius * math.sin(angle)
}
# Create edge traces
edge_traces = []
max_edge_weight = max((abs(w) for _, _, w in edges_sorted), default=0.0) or 1.0
for source_idx, target_idx, weight in edges_sorted:
x0, y0 = node_positions[source_idx]['x'], node_positions[source_idx]['y']
x1, y1 = node_positions[target_idx]['x'], node_positions[target_idx]['y']
# Color based on sign
color = '#d35400' if weight >= 0 else '#3867d6'
# Width based on magnitude
width = 0.5 + 4.5 * (abs(weight) / max_edge_weight)
edge_trace = go.Scatter(
x=[x0, x1, None],
y=[y0, y1, None],
mode='lines',
line=dict(color=color, width=width),
opacity=0.7,
hoverinfo='text',
hovertext=(
f"{_strip_occurrence_suffix(features[source_idx])} ↔ "
f"{_strip_occurrence_suffix(features[target_idx])}<br>"
f"Interaction: {weight:+.3f}"
),
showlegend=False,
)
edge_traces.append(edge_trace)
# Create node trace
node_x = []
node_y = []
node_text = []
node_colors = []
node_sizes = []
for idx in range(n):
pos = node_positions[idx]
node_x.append(pos['x'])
node_y.append(pos['y'])
# Shorten label if too long
label = _strip_occurrence_suffix(features[idx])
if len(label) > 30:
label = label[:27] + "..."
value = values[idx]
node_text.append(f"<b>{label}</b><br>Value: {value:+.3f}")
# Color based on attribution value
fill_color, _ = _token_colors(value, max_abs_value)
node_colors.append(fill_color)
# Size based on absolute value
size = 20 + 25 * (abs(value) / max_abs_value if max_abs_value > 0 else 0)
node_sizes.append(size)
node_trace = go.Scatter(
x=node_x,
y=node_y,
mode='markers+text',
marker=dict(
size=node_sizes,
color=node_colors,
line=dict(color='#5c4c78', width=2),
),
text=[f"{_strip_occurrence_suffix(features[i])[:20]}" for i in range(n)], # Short labels on nodes
textposition="top center",
textfont=dict(size=10, color='#1f1533'),
hoverinfo='text',
hovertext=node_text,
showlegend=False,
)
# Create figure
fig = go.Figure(data=edge_traces + [node_trace])
# Update layout
fig.update_layout(
title=dict(
text=f"{method.title()} pairwise interactions network",
font=dict(size=16, color='#1f1533')
),
showlegend=False,
hovermode='closest',
margin=dict(b=20, l=20, r=20, t=60),
xaxis=dict(showgrid=False, zeroline=False, showticklabels=False),
yaxis=dict(showgrid=False, zeroline=False, showticklabels=False),
plot_bgcolor='rgba(250, 248, 255, 0.5)',
paper_bgcolor='white',
height=600,
template="plotly_white",
)
# Add annotation with instructions
fig.add_annotation(
text="💡 Hover over nodes and edges to see details | Zoom and pan to explore",
xref="paper",
yref="paper",
x=0.5,
y=-0.05,
showarrow=False,
font=dict(size=11, color='#666'),
xanchor='center',
)
return fig
def _render_token_chip_view(
features: List[str],
values: List[float],
adjacency: Dict[str, List[Tuple[str, float]]],
method: str,
max_abs: float,
max_links: int,
) -> str:
def _shorten(text: str, max_len: int = 80) -> str:
text = text or ""
return text if len(text) <= max_len else text[: max_len - 1] + "…"
is_influence = method.lower() == "influence"
tokens_html: List[str] = []
for idx, token in enumerate(features):
value = abs(values[idx]) if is_influence else values[idx]
bg, border = _token_colors(value, max_abs)
display_token = _strip_occurrence_suffix(token)
label_full = escape(display_token, quote=True)
label_short = escape(_shorten(display_token), quote=True)
partners = adjacency.get(token, [])[:max_links]
partner_total = len(adjacency.get(token, []))
partner_label = "interaction" if partner_total == 1 else "interactions"
body: List[str] = []
if partners:
body.append('<ul class="token-link-list">')
for partner, val in partners:
display_partner = _strip_occurrence_suffix(str(partner))
display_val = f"{abs(val):.3f}" if is_influence else f"{val:+.3f}"
body.append(
"<li>"
f"<span class='token-link-name'>{escape(display_partner)}</span>"
f"<span class='token-link-value'>{display_val}</span>"
"</li>"
)
body.append("</ul>")
else:
body.append("<div class='token-interaction-empty'>No interactions recorded.</div>")
score_display = f"{value:.2f}" if is_influence else f"{value:+.2f}"
open_attr = " open" if idx == 0 else ""
tokens_html.append(
f'<details class="interaction-token"{open_attr} '
f'style="background-color:{bg}; border-color:{border};">'
"<summary>"
f'<span class="interaction-token__score">{score_display}</span>'
f'<span class="interaction-token__text" title="{label_full}">{label_short}</span>'
f'<span class="interaction-token__hint">{partner_total} {partner_label}</span>'
"</summary>"
+ "".join(body) +
"</details>"
)
return "".join([
_TOKEN_VIEW_STYLE,
'<div class="token-interaction-view">',
'<div class="token-interaction-panel">'
f'{escape(method.title())} pairwise interactions · click a token to inspect its strongest partners.'
"</div>",
'<div class="token-interaction-grid">',
"".join(tokens_html) or "<div class='token-interaction-empty'>No tokens available.</div>",
"</div>",
"</div>",
])
def _render_sentence_link_view(
features: List[str],
values: List[float],
adjacency: Dict[str, List[Tuple[str, float]]],
pairwise: List[Tuple[Tuple[str, ...], float]],
method: str,
row_gap: float = 70.0,
max_edges: int = 14,
) -> str:
feature_index = {feat: idx for idx, feat in enumerate(features)}
edges: List[Tuple[int, int, float]] = []
for feats, val in pairwise:
if len(feats) != 2:
continue
a_idx = feature_index.get(feats[0])
b_idx = feature_index.get(feats[1])
if a_idx is None or b_idx is None or a_idx == b_idx:
continue
edges.append((a_idx, b_idx, float(val)))
edges.sort(key=lambda item: abs(item[2]), reverse=True)
trimmed: List[Tuple[int, int, float]] = []
seen = set()
for a_idx, b_idx, val in edges:
key = tuple(sorted((a_idx, b_idx)))
if key in seen:
continue
seen.add(key)
trimmed.append((a_idx, b_idx, val))
if len(trimmed) >= max_edges:
break
max_edge = max((abs(val) for _, _, val in trimmed), default=0.0) or 1.0
canvas_height = row_gap * len(features) + 20
rows_html: List[str] = []
value_max = max((abs(v) for v in values), default=0.0) or 1.0
for idx, token in enumerate(features):
value = values[idx]
bg, border = _token_colors(value, value_max)
label_text = escape(token)
label_attr = escape(token, quote=True)
partner_badges: List[str] = []
for partner, val in adjacency.get(token, [])[:2]:
partner_badges.append(
f"<span class='sentence-row__badge'>{escape(str(partner))} {val:+.2f}</span>"
)
links_html = ""
if partner_badges:
links_html = "<div class='sentence-row__links'>Top links: " + " ".join(partner_badges) + "</div>"
rows_html.append(
f'<div class="sentence-row" style="border-color:{border};" title="{label_attr}">'
f'<span class="sentence-row__score">{value:+.2f}</span>'
f'<div class="sentence-row__text">{label_text}</div>'
f"{links_html}"
"</div>"
)
path_elements: List[str] = []
for a_idx, b_idx, val in trimmed:
y1 = row_gap * a_idx + row_gap / 2
y2 = row_gap * b_idx + row_gap / 2
color = "#d35400" if val >= 0 else "#3867d6"
width = 1.5 + 3.0 * (abs(val) / max_edge)
control = 120 + abs(a_idx - b_idx) * 12
path_elements.append(
f'<path class="sentence-link-path" d="M10 {y1:.1f} C {control:.1f} {y1:.1f}, {control + 60:.1f} {y2:.1f}, 220 {y2:.1f}" '
f'stroke="{color}" stroke-width="{width:.2f}" data-label="{escape(features[a_idx])}{escape(features[b_idx])}" />'
)
mid_y = (y1 + y2) / 2
path_elements.append(
f'<text class="sentence-link-label" x="160" y="{mid_y:.1f}">{val:+.1f}</text>'
)
node_elements = [
f'<circle class="sentence-node" cx="230" cy="{(row_gap * idx + row_gap / 2):.1f}" r="5" />'
for idx in range(len(features))
]
if path_elements:
links_block = (
f'<svg viewBox="0 0 240 {canvas_height:.0f}" preserveAspectRatio="none">'
+ "".join(path_elements + node_elements) +
"</svg>"
)
else:
links_block = "<div class='token-interaction-empty'>No pairwise arcs available.</div>"
return "".join([
_TOKEN_VIEW_STYLE,
'<div class="sentence-interaction-view">',
'<div class="sentence-interaction-header">'
f'{escape(method.title())} pairwise interactions · one sentence per row.'
"</div>",
'<div class="sentence-interaction-body">',
'<div class="sentence-list">',
"".join(rows_html),
"</div>",
f'<div class="sentence-links" style="--canvas-height:{canvas_height:.0f}px;">'
f"{links_block}</div>",
"</div>",
"</div>",
])