Download this file

226 lines (198 with data), 9.3 kB

  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
"""
Exctraction of subgraph using Prize-Collecting Steiner Tree (PCST) algorithm.
"""
from typing import Tuple, NamedTuple
import numpy as np
import torch
import pcst_fast
from torch_geometric.data.data import Data
class PCSTPruning(NamedTuple):
"""
Prize-Collecting Steiner Tree (PCST) pruning algorithm implementation inspired by G-Retriever
(He et al., 'G-Retriever: Retrieval-Augmented Generation for Textual Graph Understanding and
Question Answering', NeurIPS 2024) paper.
https://arxiv.org/abs/2402.07630
https://github.com/XiaoxinHe/G-Retriever/blob/main/src/dataset/utils/retrieval.py
Args:
topk: The number of top nodes to consider.
topk_e: The number of top edges to consider.
cost_e: The cost of the edges.
c_const: The constant value for the cost of the edges computation.
root: The root node of the subgraph, -1 for unrooted.
num_clusters: The number of clusters.
pruning: The pruning strategy to use.
verbosity_level: The verbosity level.
"""
topk: int = 3
topk_e: int = 3
cost_e: float = 0.5
c_const: float = 0.01
root: int = -1
num_clusters: int = 1
pruning: str = "gw"
verbosity_level: int = 0
def compute_prizes(self, graph: Data, query_emb: torch.Tensor) -> np.ndarray:
"""
Compute the node prizes based on the cosine similarity between the query and nodes,
as well as the edge prizes based on the cosine similarity between the query and edges.
Note that the node and edge embeddings shall use the same embedding model and dimensions
with the query.
Args:
graph: The knowledge graph in PyTorch Geometric Data format.
query_emb: The query embedding in PyTorch Tensor format.
Returns:
The prizes of the nodes and edges.
"""
# Compute prizes for nodes
n_prizes = torch.nn.CosineSimilarity(dim=-1)(query_emb, graph.x)
topk = min(self.topk, graph.num_nodes)
_, topk_n_indices = torch.topk(n_prizes, topk, largest=True)
n_prizes = torch.zeros_like(n_prizes)
n_prizes[topk_n_indices] = torch.arange(topk, 0, -1).float()
# Compute prizes for edges
# e_prizes = torch.nn.CosineSimilarity(dim=-1)(query_emb, graph.edge_attr)
# topk_e = min(self.topk_e, e_prizes.unique().size(0))
# topk_e_values, _ = torch.topk(e_prizes.unique(), topk_e, largest=True)
# e_prizes[e_prizes < topk_e_values[-1]] = 0.0
# last_topk_e_value = topk_e
# for k in range(topk_e):
# indices = e_prizes == topk_e_values[k]
# value = min((topk_e - k) / sum(indices), last_topk_e_value)
# e_prizes[indices] = value
# last_topk_e_value = value * (1 - self.c_const)
# Optimized version of the above code
e_prizes = torch.nn.CosineSimilarity(dim=-1)(query_emb, graph.edge_attr)
unique_prizes, inverse_indices = e_prizes.unique(return_inverse=True)
topk_e = min(self.topk_e, unique_prizes.size(0))
topk_e_values, _ = torch.topk(unique_prizes, topk_e, largest=True)
e_prizes[e_prizes < topk_e_values[-1]] = 0.0
last_topk_e_value = topk_e
for k in range(topk_e):
indices = inverse_indices == (
unique_prizes == topk_e_values[k]
).nonzero(as_tuple=True)[0]
value = min((topk_e - k) / indices.sum().item(), last_topk_e_value)
e_prizes[indices] = value
last_topk_e_value = value * (1 - self.c_const)
return {"nodes": n_prizes, "edges": e_prizes}
def compute_subgraph_costs(
self, graph: Data, prizes: dict
) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
"""
Compute the costs in constructing the subgraph proposed by G-Retriever paper.
Args:
graph: The knowledge graph in PyTorch Geometric Data format.
prizes: The prizes of the nodes and the edges.
Returns:
edges: The edges of the subgraph, consisting of edges and number of edges without
virtual edges.
prizes: The prizes of the subgraph.
costs: The costs of the subgraph.
"""
# Logic to reduce the cost of the edges such that at least one edge is selected
updated_cost_e = min(
self.cost_e,
prizes["edges"].max().item() * (1 - self.c_const / 2),
)
# Initialize variables
edges = []
costs = []
virtual = {
"n_prizes": [],
"edges": [],
"costs": [],
}
mapping = {"nodes": {}, "edges": {}}
# Compute the costs, edges, and virtual variables based on the prizes
for i, (src, dst) in enumerate(graph.edge_index.T.numpy()):
prize_e = prizes["edges"][i]
if prize_e <= updated_cost_e:
mapping["edges"][len(edges)] = i
edges.append((src, dst))
costs.append(updated_cost_e - prize_e)
else:
virtual_node_id = graph.num_nodes + len(virtual["n_prizes"])
mapping["nodes"][virtual_node_id] = i
virtual["edges"].append((src, virtual_node_id))
virtual["edges"].append((virtual_node_id, dst))
virtual["costs"].append(0)
virtual["costs"].append(0)
virtual["n_prizes"].append(prize_e - updated_cost_e)
prizes = np.concatenate([prizes["nodes"], np.array(virtual["n_prizes"])])
edges_dict = {}
edges_dict["edges"] = edges
edges_dict["num_prior_edges"] = len(edges)
# Final computation of the costs and edges based on the virtual costs and virtual edges
if len(virtual["costs"]) > 0:
costs = np.array(costs + virtual["costs"])
edges = np.array(edges + virtual["edges"])
edges_dict["edges"] = edges
return edges_dict, prizes, costs, mapping
def get_subgraph_nodes_edges(
self, graph: Data, vertices: np.ndarray, edges_dict: dict, mapping: dict,
) -> dict:
"""
Get the selected nodes and edges of the subgraph based on the vertices and edges computed
by the PCST algorithm.
Args:
graph: The knowledge graph in PyTorch Geometric Data format.
vertices: The vertices of the subgraph computed by the PCST algorithm.
edges_dict: The dictionary of edges of the subgraph computed by the PCST algorithm,
and the number of prior edges (without virtual edges).
mapping: The mapping dictionary of the nodes and edges.
num_prior_edges: The number of edges before adding virtual edges.
Returns:
The selected nodes and edges of the extracted subgraph.
"""
# Get edges information
edges = edges_dict["edges"]
num_prior_edges = edges_dict["num_prior_edges"]
# Retrieve the selected nodes and edges based on the given vertices and edges
subgraph_nodes = vertices[vertices < graph.num_nodes]
subgraph_edges = [mapping["edges"][e] for e in edges if e < num_prior_edges]
virtual_vertices = vertices[vertices >= graph.num_nodes]
if len(virtual_vertices) > 0:
virtual_vertices = vertices[vertices >= graph.num_nodes]
virtual_edges = [mapping["nodes"][i] for i in virtual_vertices]
subgraph_edges = np.array(subgraph_edges + virtual_edges)
edge_index = graph.edge_index[:, subgraph_edges]
subgraph_nodes = np.unique(
np.concatenate(
[subgraph_nodes, edge_index[0].numpy(), edge_index[1].numpy()]
)
)
return {"nodes": subgraph_nodes, "edges": subgraph_edges}
def extract_subgraph(self, graph: Data, query_emb: torch.Tensor) -> dict:
"""
Perform the Prize-Collecting Steiner Tree (PCST) algorithm to extract the subgraph.
Args:
graph: The knowledge graph in PyTorch Geometric Data format.
query_emb: The query embedding.
Returns:
The selected nodes and edges of the subgraph.
"""
# Assert the topk and topk_e values for subgraph retrieval
assert self.topk > 0, "topk must be greater than or equal to 0"
assert self.topk_e > 0, "topk_e must be greater than or equal to 0"
# Retrieve the top-k nodes and edges based on the query embedding
prizes = self.compute_prizes(graph, query_emb)
# Compute costs in constructing the subgraph
edges_dict, prizes, costs, mapping = self.compute_subgraph_costs(
graph, prizes
)
# Retrieve the subgraph using the PCST algorithm
result_vertices, result_edges = pcst_fast.pcst_fast(
edges_dict["edges"],
prizes,
costs,
self.root,
self.num_clusters,
self.pruning,
self.verbosity_level,
)
subgraph = self.get_subgraph_nodes_edges(
graph,
result_vertices,
{"edges": result_edges, "num_prior_edges": edges_dict["num_prior_edges"]},
mapping)
return subgraph