Python中如何将树状图转树对象并关联叶节点标签？

Hey there! Let's work through how to connect each ClusterNode (from SciPy's to_tree() method) to the positions in den['leaves'] and labels in den['ivl'] for your hierarchical text clustering. Here's a straightforward, step-by-step breakdown:

First, let's generate the root node of your cluster tree using the linkage matrix you created:

from scipy.cluster.hierarchy import to_tree

# Convert your linkage matrix `l` into a ClusterNode root object
root = to_tree(l)

The den['leaves'] list holds the original indices of your leaf nodes (matching the order in model.wv.index2word) as they appear in the dendrogram. den['ivl'] is the corresponding list of labels in that same dendrogram order. We'll create two dictionaries to bridge these values:

# Map leaf node ID → its position index in den['leaves']
leaf_id_to_pos = {leaf_id: idx for idx, leaf_id in enumerate(den['leaves'])}

# Map leaf node ID → its label from den['ivl']
# (This is the same as model.wv.index2word[leaf_id], but uses den['ivl'] directly)
leaf_id_to_label = {den['leaves'][idx]: label for idx, label in enumerate(den['ivl'])}

We can write a recursive traversal function to walk the tree, linking leaf nodes to their dendrogram positions and labels, while tracking how internal nodes merge child clusters.

Basic Depth-First Traversal (Post-Order) #

This aligns with the order clusters were merged, and matches the structure of den['leaves']:

def traverse_tree(node, id_to_pos, id_to_label):
    # Handle leaf nodes (link to dendrogram position and label)
    if node.is_leaf():
        pos = id_to_pos[node.id]
        label = id_to_label[node.id]
        print(f"Leaf Node: ID={node.id} | Dendrogram Position={pos} | Label='{label}'")
        return {
            "type": "leaf",
            "id": node.id,
            "dendrogram_pos": pos,
            "label": label
        }
    # Handle internal nodes (track merged child clusters)
    else:
        left_child = traverse_tree(node.left, id_to_pos, id_to_label)
        right_child = traverse_tree(node.right, id_to_pos, id_to_label)
        print(f"Internal Node: ID={node.id} | Merged Nodes={node.left.id} & {node.right.id}")
        return {
            "type": "internal",
            "id": node.id,
            "left_child": left_child,
            "right_child": right_child
        }

# Start traversal from the root node
tree_structure = traverse_tree(root, leaf_id_to_pos, leaf_id_to_label)

Advanced Traversal (Track All Leaves in Internal Nodes) #

If you want each internal node to include details of all leaf nodes it contains (with their positions and labels), use this modified function:

def traverse_tree_with_leaf_details(node, id_to_pos, id_to_label):
    if node.is_leaf():
        pos = id_to_pos[node.id]
        label = id_to_label[node.id]
        leaf_details = [{
            "leaf_id": node.id,
            "dendrogram_pos": pos,
            "label": label
        }]
        return {
            "type": "leaf",
            "id": node.id,
            "leaves": leaf_details
        }
    else:
        left_leaves = traverse_tree_with_leaf_details(node.left, id_to_pos, id_to_label)["leaves"]
        right_leaves = traverse_tree_with_leaf_details(node.right, id_to_pos, id_to_label)["leaves"]
        all_leaves = left_leaves + right_leaves
        print(f"Internal Node {node.id} contains {len(all_leaves)} leaves: {[leaf['label'] for leaf in all_leaves]}")
        return {
            "type": "internal",
            "id": node.id,
            "left_child_id": node.left.id,
            "right_child_id": node.right.id,
            "leaves": all_leaves
        }

tree_with_leaves = traverse_tree_with_leaf_details(root, leaf_id_to_pos, leaf_id_to_label)

• A ClusterNode's id attribute: For leaf nodes, this is the original index from your word vectors (matches model.wv.index2word). For internal nodes, IDs start at n (where n is the number of original samples) and increment as clusters are merged.
• den['leaves'] order: This is the sequence of leaf nodes as they appear in your left-oriented dendrogram (top to bottom). Our mapping dictionaries let you directly link this position to the leaf's ID and label.

内容的提问来源于stack exchange，提问作者Christina Tsangouri