File size: 17,137 Bytes
a3682cf | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 | """
models/jodie.py
===============
JODIE: Predicting Dynamic Embedding Trajectory in Temporal Interaction Networks
Kumar et al., KDD 2019
Architecture
------------
JODIE maintains dual dynamic embeddings — one per node role:
- User (sender) embedding: h_u ← updated on each outgoing event
- Item (receiver) embedding: h_v ← updated on each incoming event
Key ideas:
1. Time projection: Before each update, project the existing embedding forward
in time using a learned linear transformation conditioned on Δt:
ĥ_u(t) = (1 + W_u · Δt_emb) ⊙ h_u [element-wise time scaling]
where Δt_emb = linear(Δt) is a learnable time embedding.
2. RNN update: After projection, the RNN ingests the *other node's projected
embedding* concatenated with edge features:
h_u ← RNN( cat(ĥ_v, edge_feat), ĥ_u )
h_v ← RNN( cat(ĥ_u, edge_feat), ĥ_v )
3. Node classifier: operates on the latest projected h_u at evaluation time.
This is a faithful re-implementation of the JODIE equations from the KDD'19 paper,
adapted to the event-stream training loop of the upi-sim benchmark.
"""
from __future__ import annotations
from typing import List
import numpy as np
import pandas as pd
import torch
import torch.nn as nn
from models.base import TemporalModel
from src.graph.graph_builder import build_edge_features
# ------------------------------------------------------------------ #
# Core JODIE nn.Module #
# ------------------------------------------------------------------ #
class _JODIEModule(nn.Module):
def __init__(self, memory_dim: int, edge_dim: int, time_emb_dim: int = 16):
super().__init__()
self.memory_dim = memory_dim
# Time embedding: scalar Δt → vector
self.time_emb = nn.Linear(1, time_emb_dim)
# Projection: (1 + W · Δt_emb) ⊙ h — element-wise scale
self.W_proj_u = nn.Linear(time_emb_dim, memory_dim, bias=False)
self.W_proj_v = nn.Linear(time_emb_dim, memory_dim, bias=False)
# RNN: ingests projected other-node embedding + edge feature
self.rnn_u = nn.GRUCell(memory_dim + edge_dim, memory_dim)
self.rnn_v = nn.GRUCell(memory_dim + edge_dim, memory_dim)
# LayerNorm after GRU — critical for numerical stability with large Δt
self.norm_u = nn.LayerNorm(memory_dim)
self.norm_v = nn.LayerNorm(memory_dim)
# Node fraud classifier (applied to sender embedding)
self.classifier = nn.Sequential(
nn.Linear(memory_dim, 64),
nn.ReLU(),
nn.Linear(64, 1),
)
def project(
self,
h: torch.Tensor, # (B, mem_dim)
dt: torch.Tensor, # (B,)
W_proj: nn.Linear,
) -> torch.Tensor:
"""Time-projection: ĥ = (1 + W_proj(φ(Δt))) ⊙ h.
Clamp Δt and the scale factor to prevent explosions with large time gaps.
"""
dt_clamped = dt.clamp(0.0, 5.0) # normalised Δt bounded [0, 5]
dt_emb = torch.relu(self.time_emb(dt_clamped.unsqueeze(-1))) # (B, time_emb_dim)
scale = (1.0 + W_proj(dt_emb)).clamp(-2.0, 2.0) # (B, mem_dim)
return scale * h
def update(
self,
h_self: torch.Tensor, # (B, mem_dim) current (projected)
h_other: torch.Tensor, # (B, mem_dim) other endpoint (projected)
edge_feat: torch.Tensor, # (B, edge_dim)
rnn: nn.GRUCell,
norm: nn.LayerNorm,
) -> torch.Tensor:
inp = torch.cat([h_other, edge_feat], dim=-1)
out = rnn(inp, h_self)
return norm(out) # stabilise magnitude after GRU
def classify(self, h: torch.Tensor) -> torch.Tensor:
return self.classifier(h).squeeze(-1)
# ------------------------------------------------------------------ #
# JODIEWrapper (TemporalModel interface) #
# ------------------------------------------------------------------ #
class JODIEWrapper(TemporalModel):
"""JODIE dual-RNN temporal model with time-projection embeddings."""
def __init__(
self,
memory_dim: int = 64,
time_emb_dim: int = 16,
device: str = "cpu",
):
self.memory_dim = memory_dim
self.time_emb_dim = time_emb_dim
self.device = torch.device(device)
self._module: _JODIEModule | None = None
self._memory: torch.Tensor | None = None # (n_nodes, mem_dim)
self._last_t: torch.Tensor | None = None # (n_nodes,)
self._norm_stats: dict | None = None
self._n_nodes: int = 0
self._edge_dim: int = 0
@property
def name(self) -> str:
return "JODIE"
# ------------------------------------------------------------------ #
def fit(self, df_train: pd.DataFrame, num_epochs: int = 3) -> None:
df_train = df_train.sort_values("timestamp").reset_index(drop=True)
ef_np = build_edge_features(df_train).astype(np.float32)
edge_dim = ef_np.shape[1]
self._edge_dim = edge_dim
ea_mean = ef_np.mean(axis=0)
ea_std = ef_np.std(axis=0) + 1e-6
ef_np = (ef_np - ea_mean) / ea_std
t_vals = df_train["timestamp"].values.astype(np.float32)
t_min, t_max = t_vals.min(), t_vals.max()
t_norm = (t_vals - t_min) / (t_max - t_min + 1e-6)
self._norm_stats = {
"ea_mean": ea_mean, "ea_std": ea_std,
"t_min": t_min, "t_max": t_max,
}
all_ids = np.union1d(df_train["sender_id"].values, df_train["receiver_id"].values)
n_nodes = int(all_ids.max()) + 1
self._n_nodes = n_nodes
module = _JODIEModule(
memory_dim=self.memory_dim,
edge_dim=edge_dim,
time_emb_dim=self.time_emb_dim,
).to(self.device)
self._module = module
memory = torch.zeros(n_nodes, self.memory_dim, device=self.device)
last_t = torch.zeros(n_nodes, device=self.device)
self._memory = memory
self._last_t = last_t
u_ids = torch.tensor(df_train["sender_id"].values, dtype=torch.long)
v_ids = torch.tensor(df_train["receiver_id"].values, dtype=torch.long)
ef_all = torch.tensor(ef_np, dtype=torch.float32)
t_all = torch.tensor(t_norm, dtype=torch.float32)
y_all = torch.tensor(df_train["is_fraud"].values, dtype=torch.float32)
raw_pw = (y_all == 0).sum() / ((y_all == 1).sum() + 1e-6)
pos_weight = torch.clamp(raw_pw, max=10.0).to(self.device)
loss_fn = nn.BCEWithLogitsLoss(pos_weight=pos_weight)
# Edge-level classifier for proxy supervision during training
edge_clf = nn.Sequential(
nn.Linear(self.memory_dim * 2 + edge_dim, 64),
nn.ReLU(),
nn.Linear(64, 1),
).to(self.device)
self._edge_clf = edge_clf
opt = torch.optim.Adam(
list(module.parameters()) + list(edge_clf.parameters()),
lr=1e-3,
)
batch_size = 512
N = len(df_train)
for epoch in range(num_epochs):
memory.zero_()
last_t.zero_()
total_loss = 0.0
for i in range(0, N, batch_size):
j = min(i + batch_size, N)
u_b = u_ids[i:j].to(self.device)
v_b = v_ids[i:j].to(self.device)
t_b = t_all[i:j].to(self.device)
ef_b = ef_all[i:j].to(self.device)
y_b = y_all[i:j].to(self.device)
h_u = memory[u_b].clone()
h_v = memory[v_b].clone()
dt_u = (t_b - last_t[u_b]).clamp(min=0.0)
dt_v = (t_b - last_t[v_b]).clamp(min=0.0)
# Time projection
h_u_proj = module.project(h_u.detach(), dt_u, module.W_proj_u)
h_v_proj = module.project(h_v.detach(), dt_v, module.W_proj_v)
# JODIE update (LayerNorm inside update() for stability)
h_u_new = module.update(h_u_proj, h_v_proj.detach(), ef_b, module.rnn_u, module.norm_u)
h_v_new = module.update(h_v_proj, h_u_proj.detach(), ef_b, module.rnn_v, module.norm_v)
# Scatter-based memory write — guard against NaN
both = torch.cat([u_b, v_b])
both_h = torch.nan_to_num(torch.cat([h_u_new, h_v_new], dim=0), nan=0.0)
unique_ids, inv = torch.unique(both, return_inverse=True)
agg_h = torch.zeros(len(unique_ids), self.memory_dim, device=self.device)
agg_h.index_add_(0, inv, both_h.detach())
cnt = torch.bincount(inv).unsqueeze(1).float()
memory[unique_ids] = agg_h / cnt
last_t[u_b] = t_b
last_t[v_b] = t_b
# Loss: edge-level fraud classification
ef_concat = torch.cat([h_u_new, h_v_new, ef_b], dim=-1)
logits = edge_clf(ef_concat).squeeze(-1)
logits = torch.clamp(logits, -10, 10)
loss = loss_fn(logits, y_b)
if not torch.isnan(loss):
opt.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(module.parameters(), 1.0)
opt.step()
total_loss += loss.item()
print(f"[JODIE] Epoch {epoch + 1}/{num_epochs} Loss: {total_loss:.4f}")
# Node classifier on sender memory
self._node_clf = nn.Sequential(
nn.Linear(self.memory_dim, 64),
nn.ReLU(),
nn.Linear(64, 1),
).to(self.device)
# ------------------------------------------------------------------ #
def predict(self, df_eval: pd.DataFrame, eval_nodes: List[int]) -> np.ndarray:
assert self._module is not None, "Call fit() first."
df_eval = df_eval.sort_values("timestamp").reset_index(drop=True)
device = self.device
module = self._module
memory = self._memory
last_t = self._last_t
ns = self._norm_stats
ef_np = build_edge_features(df_eval).astype(np.float32)
ef_np = (ef_np - ns["ea_mean"]) / ns["ea_std"]
t_vals = df_eval["timestamp"].values.astype(np.float32)
t_norm = (t_vals - ns["t_min"]) / (ns["t_max"] - ns["t_min"] + 1e-6)
u_ids = torch.tensor(df_eval["sender_id"].values, dtype=torch.long)
v_ids = torch.tensor(df_eval["receiver_id"].values, dtype=torch.long)
ef_t = torch.tensor(ef_np, dtype=torch.float32)
t_t = torch.tensor(t_norm, dtype=torch.float32)
module.eval()
batch_size = 512
with torch.no_grad():
for i in range(0, len(df_eval), batch_size):
j = min(i + batch_size, len(df_eval))
u_b = u_ids[i:j].to(device)
v_b = v_ids[i:j].to(device)
t_b = t_t[i:j].to(device)
ef_b = ef_t[i:j].to(device)
h_u = memory[u_b].clone()
h_v = memory[v_b].clone()
dt_u = (t_b - last_t[u_b]).clamp(min=0.0)
dt_v = (t_b - last_t[v_b]).clamp(min=0.0)
h_u_proj = module.project(h_u, dt_u, module.W_proj_u)
h_v_proj = module.project(h_v, dt_v, module.W_proj_v)
h_u_new = module.update(h_u_proj, h_v_proj, ef_b, module.rnn_u, module.norm_u)
h_v_new = module.update(h_v_proj, h_u_proj, ef_b, module.rnn_v, module.norm_v)
both = torch.cat([u_b, v_b])
both_h = torch.nan_to_num(torch.cat([h_u_new, h_v_new], dim=0), nan=0.0)
unique_ids, inv = torch.unique(both, return_inverse=True)
agg_h = torch.zeros(len(unique_ids), self.memory_dim, device=device)
agg_h.index_add_(0, inv, both_h)
cnt = torch.bincount(inv).unsqueeze(1).float()
memory[unique_ids] = agg_h / cnt
last_t[u_b] = t_b
last_t[v_b] = t_b
eval_t = torch.tensor(
[min(n, self._n_nodes - 1) for n in eval_nodes],
dtype=torch.long, device=device,
)
node_emb = memory[eval_t]
# Guard: init classifier if train_node_classifier was never called
if not hasattr(self, "_node_clf") or self._node_clf is None:
self._node_clf = nn.Sequential(
nn.Linear(self.memory_dim, 64), nn.ReLU(), nn.Linear(64, 1)
).to(device)
with torch.no_grad():
probs = torch.sigmoid(self._node_clf(node_emb).squeeze(-1)).cpu().numpy()
return probs.astype(np.float32)
def extract_prefix_embeddings(
self,
df_eval: pd.DataFrame,
examples: pd.DataFrame,
) -> np.ndarray:
assert self._module is not None, "Call fit() first."
if examples.empty:
return np.zeros((0, self.memory_dim), dtype=np.float32)
df_eval = df_eval.sort_values("timestamp").reset_index(drop=True).copy()
if "local_event_idx" not in df_eval.columns:
df_eval["local_event_idx"] = df_eval.groupby("sender_id").cumcount().astype(np.int32)
capture_map: dict[tuple[int, int], list[int]] = {}
for ex_idx, row in enumerate(examples.itertuples(index=False)):
key = (int(row.sender_id), int(row.eval_local_event_idx))
capture_map.setdefault(key, []).append(ex_idx)
max_seen_id = int(max(df_eval["sender_id"].max(), df_eval["receiver_id"].max())) + 1
memory = torch.zeros(max(self._n_nodes, max_seen_id), self.memory_dim, device=self.device)
last_t = torch.zeros(max(self._n_nodes, max_seen_id), device=self.device)
ns = self._norm_stats
module = self._module
ef_np = build_edge_features(df_eval).astype(np.float32)
ef_np = (ef_np - ns["ea_mean"]) / ns["ea_std"]
t_vals = df_eval["timestamp"].to_numpy(dtype=np.float32)
t_norm = (t_vals - ns["t_min"]) / (ns["t_max"] - ns["t_min"] + 1e-6)
out = np.zeros((len(examples), self.memory_dim), dtype=np.float32)
module.eval()
with torch.no_grad():
for idx, row in enumerate(df_eval.itertuples(index=False)):
u = torch.tensor([int(row.sender_id)], dtype=torch.long, device=self.device)
v = torch.tensor([int(row.receiver_id)], dtype=torch.long, device=self.device)
t = torch.tensor([t_norm[idx]], dtype=torch.float32, device=self.device)
ef = torch.tensor(ef_np[idx:idx + 1], dtype=torch.float32, device=self.device)
h_u = memory[u].clone()
h_v = memory[v].clone()
dt_u = (t - last_t[u]).clamp(min=0.0)
dt_v = (t - last_t[v]).clamp(min=0.0)
h_u_proj = module.project(h_u, dt_u, module.W_proj_u)
h_v_proj = module.project(h_v, dt_v, module.W_proj_v)
h_u_new = module.update(h_u_proj, h_v_proj, ef, module.rnn_u, module.norm_u)
h_v_new = module.update(h_v_proj, h_u_proj, ef, module.rnn_v, module.norm_v)
both_ids = torch.cat([u, v])
both_h = torch.nan_to_num(torch.cat([h_u_new, h_v_new], dim=0), nan=0.0)
unique_ids, inv = torch.unique(both_ids, return_inverse=True)
agg_h = torch.zeros(len(unique_ids), self.memory_dim, device=self.device)
agg_h.index_add_(0, inv, both_h)
cnt = torch.bincount(inv).unsqueeze(1).float()
memory[unique_ids] = agg_h / cnt
last_t[u] = t
last_t[v] = t
key = (int(row.sender_id), int(row.local_event_idx))
if key in capture_map:
emb = memory[int(row.sender_id)].detach().cpu().numpy().astype(np.float32)
for ex_idx in capture_map[key]:
out[ex_idx] = emb
return out
# ------------------------------------------------------------------ #
def reset_memory(self) -> None:
if self._memory is not None:
self._memory.zero_()
self._last_t.zero_()
# ------------------------------------------------------------------ #
def train_node_classifier(
self, eval_nodes: List[int], y_labels: np.ndarray, num_epochs: int = 150
) -> None:
device = self.device
eval_t = torch.tensor(eval_nodes, dtype=torch.long, device=device)
node_emb = self._memory[eval_t].detach()
y = torch.tensor(y_labels, dtype=torch.float32, device=device)
pw = torch.clamp((y == 0).sum() / ((y == 1).sum() + 1e-6), max=10.0)
loss_fn = nn.BCEWithLogitsLoss(pos_weight=pw)
opt = torch.optim.Adam(self._node_clf.parameters(), lr=1e-3)
self._node_clf.train()
for _ in range(num_epochs):
logits = self._node_clf(node_emb).squeeze(-1)
loss = loss_fn(logits, y)
opt.zero_grad()
loss.backward()
opt.step()
self._node_clf.eval()
|