File size: 28,473 Bytes
2319f81
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
"""
MARS: Multi-scale Adaptive Recurrence with State compression
============================================================

An innovative method for super long sequence modeling in sequential recommendation.

Key innovations:
1. Temporal-Aware Delta Network (TADN) for O(n) long-range modeling
   - Explicit exponential temporal decay in state updates
   - Input-dependent gating for selective memory retention
   
2. Compressive Memory Tokens
   - Fixed-size learnable memory that compresses arbitrarily long histories
   - Acts as information bottleneck (denoising effect per Rec2PM)
   
3. Dual-Branch Architecture with Learned Fusion
   - Long-term branch: TADN layers processing full history at O(n) cost
   - Short-term branch: Standard self-attention on recent K interactions
   - Adaptive gating fusion that balances long/short-term signals per user

4. Multi-Scale Temporal Encoding
   - Absolute time embeddings + relative time deltas + periodic components
   - Captures daily/weekly/seasonal patterns in user behavior

This combines ideas from HyTRec (2602.18283), Rec2PM (2602.11605), 
SIGMA (2408.11451), and HSTU (2402.17152) into a unified architecture.
"""

import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Optional, Tuple, Dict


class TemporalEncoding(nn.Module):
    """Multi-scale temporal encoding with periodic components.
    
    Captures absolute time, relative time deltas, and periodic patterns
    (daily, weekly cycles) in user behavior.
    """
    
    def __init__(self, embed_dim: int, max_periods: int = 4):
        super().__init__()
        self.embed_dim = embed_dim
        
        # Relative time delta projection
        self.time_delta_proj = nn.Linear(1, embed_dim)
        
        # Periodic components (daily=86400s, weekly=604800s, etc.)
        periods = [3600, 86400, 604800, 2592000][:max_periods]
        self.register_buffer('periods', torch.tensor(periods, dtype=torch.float32))
        self.periodic_proj = nn.Linear(max_periods * 2, embed_dim)  # sin + cos
        
        # Learnable position encoding for sequence order
        self.layernorm = nn.LayerNorm(embed_dim)
    
    def forward(self, timestamps: torch.Tensor) -> torch.Tensor:
        """
        Args:
            timestamps: (batch, seq_len) absolute timestamps in seconds
        Returns:
            temporal_emb: (batch, seq_len, embed_dim)
        """
        B, T = timestamps.shape
        
        # 1. Relative time deltas (seconds since previous interaction)
        time_deltas = torch.zeros_like(timestamps)
        time_deltas[:, 1:] = timestamps[:, 1:] - timestamps[:, :-1]
        time_deltas = time_deltas.clamp(min=0)
        # Log-scale for better numerical properties
        log_deltas = torch.log1p(time_deltas).unsqueeze(-1)  # (B, T, 1)
        delta_emb = self.time_delta_proj(log_deltas)  # (B, T, D)
        
        # 2. Periodic components
        ts_expanded = timestamps.unsqueeze(-1)  # (B, T, 1)
        periods = self.periods.view(1, 1, -1)   # (1, 1, P)
        angles = 2 * math.pi * ts_expanded / periods  # (B, T, P)
        periodic_features = torch.cat([
            torch.sin(angles),
            torch.cos(angles)
        ], dim=-1)  # (B, T, 2*P)
        periodic_emb = self.periodic_proj(periodic_features)  # (B, T, D)
        
        # 3. Combine
        temporal_emb = self.layernorm(delta_emb + periodic_emb)
        return temporal_emb


class TADNLayer(nn.Module):
    """Temporal-Aware Delta Network Layer.
    
    Linear complexity O(n) recurrent layer with:
    - Delta rule state updates (inspired by HyTRec)
    - Explicit temporal decay gating
    - Input-dependent selective memory
    
    The state matrix S is updated as:
        S_t = S_{t-1} * (I - g_t * beta_t * k_t * k_t^T) + beta_t * v_t * k_t^T
    
    where g_t incorporates temporal decay:
        g_t = alpha * sigmoid(W_g * [h_t, delta_h_t]) * tau_t + (1-alpha) * g_static
        tau_t = exp(-(t_current - t_behavior) / T)
    """
    
    def __init__(self, embed_dim: int, state_dim: int = 64, dropout: float = 0.1):
        super().__init__()
        self.embed_dim = embed_dim
        self.state_dim = state_dim
        
        # Query, Key, Value projections
        self.q_proj = nn.Linear(embed_dim, state_dim)
        self.k_proj = nn.Linear(embed_dim, state_dim)
        self.v_proj = nn.Linear(embed_dim, embed_dim)
        
        # Gating mechanism
        self.gate_proj = nn.Linear(embed_dim * 2, embed_dim)
        self.beta_proj = nn.Linear(embed_dim, state_dim)
        
        # Temporal decay parameters
        self.alpha = nn.Parameter(torch.tensor(0.5))
        self.time_scale = nn.Parameter(torch.tensor(1.0))
        
        # Static gate (learnable baseline)
        self.gate_static = nn.Parameter(torch.ones(embed_dim) * 0.5)
        
        # Output
        self.out_proj = nn.Linear(embed_dim, embed_dim)
        self.layernorm = nn.LayerNorm(embed_dim)
        self.dropout = nn.Dropout(dropout)
    
    def forward(
        self,
        x: torch.Tensor,
        timestamps: Optional[torch.Tensor] = None,
        mask: Optional[torch.Tensor] = None
    ) -> torch.Tensor:
        """
        Args:
            x: (batch, seq_len, embed_dim) input sequence
            timestamps: (batch, seq_len) timestamps for temporal decay
            mask: (batch, seq_len) boolean mask (True = valid)
        Returns:
            output: (batch, seq_len, embed_dim)
        """
        B, T, D = x.shape
        
        # Project to Q, K, V
        q = self.q_proj(x)   # (B, T, state_dim)
        k = self.k_proj(x)   # (B, T, state_dim)
        v = self.v_proj(x)   # (B, T, D)
        
        # Beta (key importance scaling)
        beta = torch.sigmoid(self.beta_proj(x))  # (B, T, state_dim)
        
        # Temporal decay
        if timestamps is not None:
            # Compute recency-based decay with proper normalization
            # Use the LAST VALID position's timestamp as reference
            # Normalize by log(1 + delta) to handle large time ranges (seconds → years)
            t_last = timestamps[:, -1:].unsqueeze(-1)  # (B, 1, 1) - last timestamp
            t_behavior = timestamps.unsqueeze(-1)       # (B, T, 1)
            time_delta = (t_last - t_behavior).clamp(min=0)
            
            # Log-normalize: log(1 + delta_seconds / 3600) → hours-scale
            log_delta = torch.log1p(time_delta / 3600.0)  # Normalize to hours
            
            # Learnable time scale controls the decay rate
            tau = torch.exp(
                -log_delta / (torch.abs(self.time_scale) * 10.0 + 1.0)
            )  # (B, T, 1), values in [0, 1]
        else:
            # Fallback: linear decay
            positions = torch.arange(T, device=x.device).float()
            tau = torch.exp(-positions / (T + 1e-6)).view(1, T, 1)
        
        # Dynamic gating with temporal awareness
        # Delta of hidden states for change detection
        x_shifted = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1)
        delta_x = x - x_shifted
        gate_input = torch.cat([x, delta_x], dim=-1)  # (B, T, 2*D)
        
        alpha = torch.sigmoid(self.alpha)
        g_dynamic = torch.sigmoid(self.gate_proj(gate_input))  # (B, T, D)
        g = alpha * g_dynamic * tau + (1 - alpha) * torch.sigmoid(self.gate_static)
        
        # Recurrent state update with delta rule
        # Use chunked processing for better GPU utilization
        chunk_size = min(64, T)  # Process in chunks for efficiency
        
        outputs = []
        S = torch.zeros(B, self.state_dim, D, device=x.device)  # State matrix
        
        for chunk_start in range(0, T, chunk_size):
            chunk_end = min(chunk_start + chunk_size, T)
            
            for t in range(chunk_start, chunk_end):
                k_t = k[:, t]     # (B, state_dim)
                v_t = v[:, t]     # (B, D)
                beta_t = beta[:, t]  # (B, state_dim)
                g_t = g[:, t]     # (B, D)
                q_t = q[:, t]     # (B, state_dim)
                
                # Delta rule update: erase old, write new
                # Clamp erase to [0, 1] for stability
                erase = torch.einsum('bs,bd->bsd', beta_t * k_t, g_t).clamp(0, 1)
                write = torch.einsum('bs,bd->bsd', beta_t * k_t, v_t)
                
                if mask is not None:
                    valid = mask[:, t].float().view(B, 1, 1)
                    S = S * (1 - erase * valid) + write * valid
                else:
                    S = S * (1 - erase) + write
                
                # Clamp state for numerical stability
                S = S.clamp(-10, 10)
                
                # Read from state
                out_t = torch.einsum('bs,bsd->bd', q_t, S)
                outputs.append(out_t)
        
        output = torch.stack(outputs, dim=1)  # (B, T, D)
        output = self.out_proj(self.dropout(output))
        output = self.layernorm(x + output)  # Residual connection
        
        return output


class CompressiveMemory(nn.Module):
    """Compressive Memory Module.
    
    Compresses long sequence history into a fixed number of memory tokens.
    Acts as information bottleneck (denoising per Rec2PM theory).
    
    Uses cross-attention: memory queries attend to sequence to extract summary.
    """
    
    def __init__(self, embed_dim: int, num_memory_tokens: int = 8, num_heads: int = 2):
        super().__init__()
        self.num_memory_tokens = num_memory_tokens
        
        # Learnable memory query tokens
        self.memory_queries = nn.Parameter(
            torch.randn(num_memory_tokens, embed_dim) * 0.02
        )
        
        # Cross-attention: memory queries attend to sequence
        self.cross_attn = nn.MultiheadAttention(
            embed_dim=embed_dim,
            num_heads=num_heads,
            batch_first=True,
            dropout=0.1
        )
        
        self.ffn = nn.Sequential(
            nn.Linear(embed_dim, embed_dim * 4),
            nn.GELU(),
            nn.Dropout(0.1),
            nn.Linear(embed_dim * 4, embed_dim),
            nn.Dropout(0.1),
        )
        
        self.norm1 = nn.LayerNorm(embed_dim)
        self.norm2 = nn.LayerNorm(embed_dim)
    
    def forward(
        self,
        sequence: torch.Tensor,
        mask: Optional[torch.Tensor] = None
    ) -> torch.Tensor:
        """
        Args:
            sequence: (batch, seq_len, embed_dim) - encoded sequence
            mask: (batch, seq_len) boolean mask (True = valid, False = padding)
        Returns:
            memory: (batch, num_memory_tokens, embed_dim)
        """
        B = sequence.shape[0]
        
        # Expand memory queries for batch
        queries = self.memory_queries.unsqueeze(0).expand(B, -1, -1)  # (B, M, D)
        
        # Cross-attention with key padding mask
        # nn.MultiheadAttention expects key_padding_mask where True = IGNORE
        if mask is not None:
            key_padding_mask = ~mask  # Invert: True means padding (to ignore)
        else:
            key_padding_mask = None
        
        attn_out, _ = self.cross_attn(
            query=queries,
            key=sequence,
            value=sequence,
            key_padding_mask=key_padding_mask
        )
        memory = self.norm1(queries + attn_out)
        memory = self.norm2(memory + self.ffn(memory))
        
        return memory


class ShortTermAttention(nn.Module):
    """Standard self-attention block for short-term (recent) interactions.
    
    Uses standard causal multi-head attention — full expressiveness
    for the most recent K items where O(K²) is acceptable.
    """
    
    def __init__(self, embed_dim: int, num_heads: int = 2, num_layers: int = 2, dropout: float = 0.1):
        super().__init__()
        encoder_layer = nn.TransformerEncoderLayer(
            d_model=embed_dim,
            nhead=num_heads,
            dim_feedforward=embed_dim * 4,
            dropout=dropout,
            activation='gelu',
            batch_first=True,
            norm_first=True
        )
        self.encoder = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)
    
    def forward(
        self,
        x: torch.Tensor,
        mask: Optional[torch.Tensor] = None
    ) -> torch.Tensor:
        """
        Args:
            x: (batch, K, embed_dim) recent interactions
            mask: (batch, K) boolean mask
        Returns:
            output: (batch, K, embed_dim)
        """
        T = x.shape[1]
        
        # Causal mask
        causal_mask = torch.triu(
            torch.ones(T, T, device=x.device, dtype=torch.bool), diagonal=1
        )
        
        # Padding mask
        src_key_padding_mask = ~mask if mask is not None else None
        
        output = self.encoder(
            x,
            mask=causal_mask,
            src_key_padding_mask=src_key_padding_mask
        )
        return output


class AdaptiveFusionGate(nn.Module):
    """Adaptive fusion gate that balances long-term and short-term signals.
    
    Per-user, per-timestep gating:
        output = sigma(gate) * long_term + (1 - sigma(gate)) * short_term
    """
    
    def __init__(self, embed_dim: int):
        super().__init__()
        self.gate = nn.Sequential(
            nn.Linear(embed_dim * 3, embed_dim),
            nn.GELU(),
            nn.Linear(embed_dim, embed_dim),
            nn.Sigmoid()
        )
    
    def forward(
        self,
        long_term: torch.Tensor,
        short_term: torch.Tensor,
        memory: torch.Tensor
    ) -> torch.Tensor:
        """
        Args:
            long_term: (batch, embed_dim)
            short_term: (batch, embed_dim) 
            memory: (batch, embed_dim) compressed memory summary
        Returns:
            fused: (batch, embed_dim)
        """
        gate_input = torch.cat([long_term, short_term, memory], dim=-1)
        g = self.gate(gate_input)
        return g * long_term + (1 - g) * short_term


class MARS(nn.Module):
    """
    MARS: Multi-scale Adaptive Recurrence with State compression
    
    Architecture:
        Input: Full user interaction sequence + timestamps
            |
            v
        [Item Embedding + Temporal Encoding]
            |
            +---- Long-term Branch (TADN layers, O(n))
            |         |
            |     [Compressive Memory] → memory tokens
            |         |
            +---- Short-term Branch (Self-Attention on recent K items)
            |
            v
        [Adaptive Fusion Gate]
            |
            v
        [Prediction Head] → next item scores
    
    Args:
        num_items: number of unique items
        embed_dim: embedding dimension
        max_seq_len: maximum sequence length (can be very long, e.g. 2048)
        short_term_len: number of recent items for short-term branch
        num_memory_tokens: number of compressive memory tokens
        num_tadn_layers: number of TADN layers in long-term branch
        num_attn_layers: number of attention layers in short-term branch
        num_heads: number of attention heads
        state_dim: state dimension for TADN
        dropout: dropout rate
    """
    
    def __init__(
        self,
        num_items: int,
        embed_dim: int = 64,
        max_seq_len: int = 512,
        short_term_len: int = 50,
        num_memory_tokens: int = 8,
        num_tadn_layers: int = 3,
        num_attn_layers: int = 2,
        num_heads: int = 2,
        state_dim: int = 64,
        dropout: float = 0.1,
    ):
        super().__init__()
        self.num_items = num_items
        self.embed_dim = embed_dim
        self.max_seq_len = max_seq_len
        self.short_term_len = short_term_len
        self.num_memory_tokens = num_memory_tokens
        
        # Item embeddings (0 = padding)
        self.item_embedding = nn.Embedding(num_items + 1, embed_dim, padding_idx=0)
        
        # Temporal encoding
        self.temporal_encoding = TemporalEncoding(embed_dim)
        
        # Learnable position encoding (for short-term branch)
        self.position_embedding = nn.Embedding(max_seq_len, embed_dim)
        
        # Input processing
        self.input_norm = nn.LayerNorm(embed_dim)
        self.input_dropout = nn.Dropout(dropout)
        
        # Long-term branch: stack of TADN layers
        self.tadn_layers = nn.ModuleList([
            TADNLayer(embed_dim, state_dim, dropout)
            for _ in range(num_tadn_layers)
        ])
        
        # Compressive memory
        self.compressive_memory = CompressiveMemory(
            embed_dim, num_memory_tokens, num_heads
        )
        
        # Short-term branch: standard self-attention
        self.short_term_attn = ShortTermAttention(
            embed_dim, num_heads, num_attn_layers, dropout
        )
        
        # Adaptive fusion
        self.fusion_gate = AdaptiveFusionGate(embed_dim)
        
        # Output projection
        self.output_norm = nn.LayerNorm(embed_dim)
        self.output_proj = nn.Linear(embed_dim, embed_dim)
        
        # Initialize weights
        self._init_weights()
    
    def _init_weights(self):
        """Initialize with truncated normal distribution."""
        for name, param in self.named_parameters():
            if 'weight' in name and param.dim() >= 2:
                nn.init.trunc_normal_(param, std=0.02)
            elif 'bias' in name:
                nn.init.zeros_(param)
        
        # Special init for item embeddings
        nn.init.trunc_normal_(self.item_embedding.weight, std=0.02)
        nn.init.zeros_(self.item_embedding.weight[0])  # Padding = zero
    
    @property
    def item_embeddings(self):
        """Access item embedding table (for evaluation)."""
        return self.item_embedding
    
    def encode(
        self,
        item_ids: torch.Tensor,
        timestamps: Optional[torch.Tensor] = None,
        mask: Optional[torch.Tensor] = None
    ) -> torch.Tensor:
        """
        Encode a full sequence into user representations.
        
        Args:
            item_ids: (batch, seq_len) item indices (0 = padding)
            timestamps: (batch, seq_len) timestamps in seconds
            mask: (batch, seq_len) boolean mask (True = valid)
        Returns:
            user_emb: (batch, embed_dim) final user representation
        """
        B, T = item_ids.shape
        
        # Create mask from padding if not provided
        if mask is None:
            mask = (item_ids != 0)
        
        # 1. Item + Temporal Embeddings
        item_emb = self.item_embedding(item_ids)  # (B, T, D)
        
        if timestamps is not None:
            temp_emb = self.temporal_encoding(timestamps.float())
            item_emb = item_emb + temp_emb
        
        # Add position embeddings (only for the sequence order)
        positions = torch.arange(T, device=item_ids.device).unsqueeze(0)
        positions = positions.clamp(max=self.max_seq_len - 1)
        pos_emb = self.position_embedding(positions)
        
        item_emb = self.input_norm(item_emb + pos_emb)
        item_emb = self.input_dropout(item_emb)
        
        # 2. Long-term Branch: TADN over full sequence
        long_term_repr = item_emb
        for tadn in self.tadn_layers:
            long_term_repr = tadn(long_term_repr, timestamps, mask)
        
        # Compress long-term into memory tokens
        memory = self.compressive_memory(long_term_repr, mask)  # (B, M, D)
        memory_summary = memory.mean(dim=1)  # (B, D) - aggregated memory
        
        # Get last valid long-term representation
        # Use mask to find last valid position
        lengths = mask.sum(dim=1).long()  # (B,)
        long_term_last = long_term_repr[
            torch.arange(B, device=item_ids.device),
            (lengths - 1).clamp(min=0)
        ]  # (B, D)
        
        # 3. Short-term Branch: Attention on recent K items
        # With right-padding, valid items are at positions 0...(length-1)
        # Extract last K valid items per user
        K = min(self.short_term_len, T)
        
        # For each user, get the last K valid positions
        short_item_ids_list = []
        short_ts_list = []
        short_mask_list = []
        
        for b in range(B):
            seq_len = lengths[b].item()
            actual_k = min(K, seq_len)
            start = max(0, seq_len - K)
            end = seq_len
            
            # Extract valid items and pad to K
            ids = item_ids[b, start:end]
            pad_len = K - actual_k
            if pad_len > 0:
                ids = torch.cat([ids, torch.zeros(pad_len, dtype=ids.dtype, device=ids.device)])
            short_item_ids_list.append(ids)
            
            if timestamps is not None:
                ts = timestamps[b, start:end]
                if pad_len > 0:
                    ts = torch.cat([ts, torch.zeros(pad_len, dtype=ts.dtype, device=ts.device)])
                short_ts_list.append(ts)
            
            m = torch.zeros(K, dtype=torch.bool, device=item_ids.device)
            m[:actual_k] = True
            short_mask_list.append(m)
        
        short_item_ids = torch.stack(short_item_ids_list)  # (B, K)
        short_mask = torch.stack(short_mask_list)            # (B, K)
        
        short_emb = self.item_embedding(short_item_ids)
        
        if timestamps is not None:
            short_ts = torch.stack(short_ts_list)  # (B, K)
            short_temp = self.temporal_encoding(short_ts.float())
            short_emb = short_emb + short_temp
        
        short_positions = torch.arange(K, device=item_ids.device).unsqueeze(0)
        short_positions = short_positions.clamp(max=self.max_seq_len - 1)
        short_emb = short_emb + self.position_embedding(short_positions)
        short_emb = self.input_norm(short_emb)
        
        short_term_repr = self.short_term_attn(short_emb, short_mask)
        
        # Get last valid short-term representation
        short_lengths = short_mask.sum(dim=1).long()
        short_term_last = short_term_repr[
            torch.arange(B, device=item_ids.device),
            (short_lengths - 1).clamp(min=0)
        ]  # (B, D)
        
        # 4. Adaptive Fusion
        user_emb = self.fusion_gate(long_term_last, short_term_last, memory_summary)
        user_emb = self.output_proj(self.output_norm(user_emb))
        
        return user_emb
    
    def forward(
        self,
        batch: Dict[str, torch.Tensor]
    ) -> torch.Tensor:
        """
        Training forward pass.
        
        Expected batch format (flat tensors, matching Yambda convention):
            - item_ids: (batch, max_seq_len) padded item sequences
            - timestamps: (batch, max_seq_len) padded timestamps
            - mask: (batch, max_seq_len) boolean mask
            - positive_ids: (batch,) positive next items
            - negative_ids: (batch, num_neg) negative items
        
        Returns:
            loss: scalar BCE loss
        """
        if self.training:
            return self._training_forward(batch)
        else:
            return self._eval_forward(batch)
    
    def _training_forward(self, batch: Dict[str, torch.Tensor]) -> torch.Tensor:
        """Compute training loss with next-item prediction."""
        item_ids = batch['item_ids']       # (B, T)
        timestamps = batch.get('timestamps')  # (B, T) or None
        mask = batch.get('mask')           # (B, T)
        pos_ids = batch['positive_ids']    # (B,)
        neg_ids = batch['negative_ids']    # (B, num_neg)
        
        # Encode user sequence
        user_emb = self.encode(item_ids, timestamps, mask)  # (B, D)
        
        # Score positive and negative items
        pos_emb = self.item_embedding(pos_ids)   # (B, D)
        neg_emb = self.item_embedding(neg_ids)   # (B, num_neg, D)
        
        pos_scores = (user_emb * pos_emb).sum(dim=-1)  # (B,)
        neg_scores = torch.einsum('bd,bnd->bn', user_emb, neg_emb)  # (B, num_neg)
        
        # BPR-style loss + BCE
        pos_labels = torch.ones_like(pos_scores)
        neg_labels = torch.zeros_like(neg_scores)
        
        loss_pos = F.binary_cross_entropy_with_logits(pos_scores, pos_labels)
        loss_neg = F.binary_cross_entropy_with_logits(neg_scores, neg_labels)
        
        loss = loss_pos + loss_neg
        return loss
    
    def _eval_forward(self, batch: Dict[str, torch.Tensor]) -> torch.Tensor:
        """Eval forward: returns user embeddings."""
        item_ids = batch['item_ids']
        timestamps = batch.get('timestamps')
        mask = batch.get('mask')
        
        user_emb = self.encode(item_ids, timestamps, mask)
        return user_emb


class SASRecBaseline(nn.Module):
    """
    Standard SASRec baseline for comparison.
    Uses causal self-attention (O(n²) complexity).
    """
    
    def __init__(
        self,
        num_items: int,
        embed_dim: int = 64,
        max_seq_len: int = 200,
        num_heads: int = 2,
        num_layers: int = 2,
        dropout: float = 0.1,
    ):
        super().__init__()
        self.num_items = num_items
        self.embed_dim = embed_dim
        self.max_seq_len = max_seq_len
        
        self.item_embedding = nn.Embedding(num_items + 1, embed_dim, padding_idx=0)
        self.position_embedding = nn.Embedding(max_seq_len, embed_dim)
        
        self.input_norm = nn.LayerNorm(embed_dim)
        self.input_dropout = nn.Dropout(dropout)
        
        encoder_layer = nn.TransformerEncoderLayer(
            d_model=embed_dim,
            nhead=num_heads,
            dim_feedforward=embed_dim * 4,
            dropout=dropout,
            activation='gelu',
            batch_first=True,
            norm_first=True
        )
        self.encoder = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)
        
        self.output_norm = nn.LayerNorm(embed_dim)
        
        self._init_weights()
    
    def _init_weights(self):
        for name, param in self.named_parameters():
            if 'weight' in name and param.dim() >= 2:
                nn.init.trunc_normal_(param, std=0.02)
            elif 'bias' in name:
                nn.init.zeros_(param)
        nn.init.zeros_(self.item_embedding.weight[0])
    
    @property
    def item_embeddings(self):
        return self.item_embedding
    
    def encode(self, item_ids, timestamps=None, mask=None):
        B, T = item_ids.shape
        if mask is None:
            mask = (item_ids != 0)
        
        item_emb = self.item_embedding(item_ids)
        positions = torch.arange(T, device=item_ids.device).unsqueeze(0)
        positions = positions.clamp(max=self.max_seq_len - 1)
        item_emb = item_emb + self.position_embedding(positions)
        item_emb = self.input_norm(item_emb)
        item_emb = self.input_dropout(item_emb)
        
        causal_mask = torch.triu(torch.ones(T, T, device=item_ids.device, dtype=torch.bool), diagonal=1)
        src_key_padding_mask = ~mask
        
        output = self.encoder(item_emb, mask=causal_mask, src_key_padding_mask=src_key_padding_mask)
        
        lengths = mask.sum(dim=1).long()
        user_emb = output[torch.arange(B, device=item_ids.device), (lengths - 1).clamp(min=0)]
        user_emb = self.output_norm(user_emb)
        
        return user_emb
    
    def forward(self, batch):
        if self.training:
            item_ids = batch['item_ids']
            timestamps = batch.get('timestamps')
            mask = batch.get('mask')
            pos_ids = batch['positive_ids']
            neg_ids = batch['negative_ids']
            
            user_emb = self.encode(item_ids, timestamps, mask)
            pos_emb = self.item_embedding(pos_ids)
            neg_emb = self.item_embedding(neg_ids)
            
            pos_scores = (user_emb * pos_emb).sum(dim=-1)
            neg_scores = torch.einsum('bd,bnd->bn', user_emb, neg_emb)
            
            loss_pos = F.binary_cross_entropy_with_logits(pos_scores, torch.ones_like(pos_scores))
            loss_neg = F.binary_cross_entropy_with_logits(neg_scores, torch.zeros_like(neg_scores))
            
            return loss_pos + loss_neg
        else:
            return self.encode(batch['item_ids'], batch.get('timestamps'), batch.get('mask'))