如何使用graph-tool将hSBM生成的分层二分图转换为仅含词与词组的圆形图？

Hey there! Let’s break down how to fix your visualization issue and get that circular word-topic graph you’re after. First, let’s clear up why your previous attempts didn’t work, then walk through the correct approach step by step.

Why Your Earlier Methods Failed #

Let’s quickly diagnose the problems with your two tries:

1. Anchor Node Approach: Replacing documents with a single anchor creates a star-shaped graph (all words pointing to the center), which doesn’t capture thematic relationships between words. You also lose the original hSBM topic block metadata when deleting document nodes.
2. Manual Edge Addition & Projection: Your filter vfilt=model.g.vp.kind.a == 0 selects document nodes instead of word nodes, so you ended up projecting a document graph instead of a word graph. Plus, manually adding edges between neighbors is inefficient and doesn’t preserve meaningful co-occurrence weights.

The Correct Workflow #

We’ll leverage graph-tool’s built-in tools to project the bipartite graph to words only, retain the hSBM topic assignments, and then visualize with a circular layout grouped by topics.

Step 1: Project the Bipartite Graph to a Word Co-Occurrence Graph

graph-tool has a project_graph function designed exactly for this—it creates a graph of word nodes where edges represent how often two words appear together in documents.

from graph_tool.all import *

# Project the original bipartite graph (model.g) to word nodes (kind=1)
# This creates a graph where edges have a "weight" property equal to co-occurrence count
word_graph = project_graph(model.g, model.g.vp['kind'], target=1)

Step 2: Migrate hSBM Topic Block Metadata

Your trained hSBM already has topic assignments for word nodes. We need to copy this metadata to our new word graph. Assuming your original model stores word topics in a vertex property like membership:

# Extract word-to-topic mapping from the original bipartite graph
original_word_nodes = [v for v in model.g.vertices() if model.g.vp['kind'][v] == 1]
word_to_topic = {
    model.g.vp['name'][v]: model.g.vp['membership'][v] 
    for v in original_word_nodes
}

# Add topic property to the new word graph
word_graph.vp['topic'] = word_graph.new_vp("int")
for v in word_graph.vertices():
    word_graph.vp['topic'][v] = word_to_topic[word_graph.vp['name'][v]]

Step 3: Generate Circular Layout Grouped by Topic

To get a clean circular layout where words from the same topic are clustered together, sort the nodes by their topic before applying the circular layout:

# Sort vertices by their topic to group them in the layout
sorted_vertices = sorted(word_graph.vertices(), key=lambda v: word_graph.vp['topic'][v])

# Create circular layout using the sorted node order
pos = circular_layout(word_graph, order=sorted_vertices)

# Assign unique colors to each topic for clarity
num_topics = len(set(word_graph.vp['topic'].a))
topic_colors = random_colors(num_topics)
vertex_fill_color = word_graph.new_vp("vector<float>")
for v in word_graph.vertices():
    vertex_fill_color[v] = topic_colors[word_graph.vp['topic'][v]]

# Draw the graph
graph_draw(
    word_graph,
    pos=pos,
    vertex_fill_color=vertex_fill_color,
    vertex_text=word_graph.vp['name'],
    vertex_font_size=9,
    edge_pen_width=word_graph.ep['weight'] / 15,  # Scale edge width by co-occurrence count
    output_size=(1200, 1200),
    output="word_topic_circular.png"
)

If You Don’t Need Co-Occurrence Edges #

If you just want to visualize words grouped by topic (no edges between words), you can simplify further:

# Create a minimal graph with only word nodes
simple_word_graph = Graph()
simple_word_graph.add_vertex(len(original_word_nodes))

# Add name and topic properties
simple_word_graph.vp['name'] = simple_word_graph.new_vp("string", vals=[model.g.vp['name'][v] for v in original_word_nodes])
simple_word_graph.vp['topic'] = simple_word_graph.new_vp("int", vals=[model.g.vp['membership'][v] for v in original_word_nodes])

# Generate sorted circular layout and draw
sorted_v = sorted(simple_word_graph.vertices(), key=lambda v: simple_word_graph.vp['topic'][v])
pos = circular_layout(simple_word_graph, order=sorted_v)

graph_draw(
    simple_word_graph,
    pos=pos,
    vertex_fill_color=vertex_fill_color,
    vertex_text=simple_word_graph.vp['name'],
    vertex_font_size=10,
    output="word_topic_simple_circle.png"
)

Key Notes #

• You don’t need to retrain the hSBM model—we’re just repurposing the topic assignments it already generated.
• Using project_graph is far more efficient than manual edge creation, and it preserves meaningful co-occurrence weights for edges.

内容的提问来源于stack exchange，提问作者Jakob Rasmussen