# Behavior Sequence Transformer

> Alibaba 2019 BST — user behavior sequence + target → Transformer encoder + concat with user features → MLP → CTR

Pick for sequential CTR when user behavior history matters (session intent). Production-deployed at Alibaba scale; needs sequence-aware features.

- Category: Recommendation
- Layers: 20
- Parameters: 128.76M
- Input shape (batchless): 50
- Output shape: 1
- Verifier verdict: pass
- Graph JSON: https://neurarch.com/templates/bst/model.json
- Open on the canvas: https://neurarch.com/?template=bst

## Structure

| # | Layer | Type | Parameters | Output shape |
| --- | --- | --- | --- | --- |
| 1 | Behavior Seq (T items) | Input | shape=[50] | 50 |
| 2 | Item Embed | Embedding | vocabSize=1000000 | 50 × 64 |
| 3 | Pos Encoding | Positional Encoding | maxLen=50 | 50 × 64 |
| 4 | Target Item | Input | shape=[1] | 1 |
| 5 | Target Embed | Embedding | vocabSize=1000000 | 1 × 64 |
| 6 | [seq; tgt] | Concatenate |  | 51 × 64 |
| 7 | Transformer Encoder | Transformer Block | numHeads=8, numLayers=1 | 51 × 64 |
| 8 | to_channels | Permute |  | 64 × 51 |
| 9 | Mean Pool | GlobalAvgPool1D |  | 64 |
| 10 | User Profile | Input | shape=[16] | 16 |
| 11 | User Tower | Linear | outFeatures=32, inFeatures=16 | 32 |
| 12 | Concat all | Concatenate |  | 96 |
| 13 | MLP 1024 | Linear | outFeatures=1024, inFeatures=96 | 1024 |
| 14 | PReLU | PReLU |  | 1024 |
| 15 | MLP 512 | Linear | outFeatures=512, inFeatures=1024 | 512 |
| 16 | PReLU | PReLU |  | 512 |
| 17 | MLP 256 | Linear | outFeatures=256, inFeatures=512 | 256 |
| 18 | CTR Head | Linear | outFeatures=1, inFeatures=256 | 1 |
| 19 | Sigmoid | Sigmoid |  | 1 |
| 20 | P(click) | Output |  | 1 |

## Verifier findings

- **info** `output-activation` at `Sigmoid`: "Sigmoid" feeds directly into Output. PyTorch's nn.CrossEntropyLoss already applies log-softmax internally, an explicit Softmax causes double-application and degrades training stability. Fix: Remove Softmax/Sigmoid for training. Restore it in a separate inference wrapper or ONNX export.
- **info** `vanishing-gradient` at `Sigmoid`: Sigmoid saturates to [0,1] / [-1,1], and its gradient approaches zero for large inputs. In networks deeper than 5 layers, this halts learning in early layers. Fix: Use ReLU, GELU, or SiLU for hidden layers. Keep Sigmoid only at binary classification outputs; Tanh in specific contexts (GAN generators, LSTM gates).
- **info** `deep-no-norm`: 16 layers with no BatchNorm, LayerNorm, or GroupNorm. Without normalization, activations can explode or vanish across layers, causing slow or unstable training. Fix: Add BatchNorm after Conv2d (CV tasks), LayerNorm after attention/FFN (NLP/LLM), or GroupNorm for small batch sizes.
- **info** `consecutive-linear-no-activation` at `MLP 256`: "MLP 256" feeds directly into "CTR Head" with no activation between them. Two stacked linear maps collapse into one (W₂·W₁), so the extra layer costs parameters but adds no representational power. Fix: Add a non-linearity (ReLU/GELU) between them. If this is a deliberate low-rank / factorized projection (down-proj → up-proj), this hint is safe to ignore.
- **info** `init-activation-mismatch` at `CTR Head`: PyTorch initializes Linear/Conv with Kaiming (He) init, which is derived for ReLU-family activations. Feeding a saturating activation (sigmoid/tanh) from a He-initialized layer starts training in the saturated tails, shrinking early gradients. Fix: Initialize these layers with Xavier instead: nn.init.xavier_uniform_(w, gain=nn.init.calculate_gain("sigmoid"|"tanh")), or switch the activation to a ReLU-family one.

## Exported PyTorch (first 46 lines)

```python
# Architecture designed with Neurarch: https://neurarch.com
# PyTorch: compatible with Python 3.8+ and torch>=1.12
# Colab: pip install torch torchvision  (usually pre-installed)
#
# WARNING: 2 layer(s) below are not yet supported by the PyTorch
# exporter and pass their input through UNCHANGED in forward():
#   - to_channels (permute)
#   - Mean Pool (globalAvgPool1d)

import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Tuple

class BehaviorSequenceTransformer(nn.Module):
    def __init__(self):
        super().__init__()

        self.embedding_1 = nn.Embedding(1000000, 64)
        self.embedding_2 = nn.Embedding(1000000, 64)
        self.transformerBlock_1 = nn.TransformerEncoderLayer(d_model=512, nhead=8, dim_feedforward=2048, batch_first=True)
        self.linear_1 = nn.Linear(16, 32)
        self.linear_2 = nn.Linear(96, 1024)
        self.prelu_1 = nn.PReLU(num_parameters=1)
        self.linear_3 = nn.Linear(1024, 512)
        self.prelu_2 = nn.PReLU(num_parameters=1)
        self.linear_4 = nn.Linear(512, 256)
        self.linear_5 = nn.Linear(256, 1)

    def forward(self, src, tgt=None):
        # Behavior Seq (T items) shape: [50]
        # Target Item shape: [1]
        # User Profile shape: [16]
        embedding_eq_emb = self.embedding_1(src)
        # positionalEncoding: add positional encoding externally (e.g. sinusoidal or learned PE)
        embedding_gt_emb = self.embedding_2(tgt)
        concatenate_at_seq = torch.cat([embedding_eq_emb, embedding_gt_emb], dim=0)
        transformer_block_trf = self.transformerBlock_1(concatenate_at_seq)
        # TODO: layer 'to_channels' (permute) is not yet supported by the exporter; passing through unchanged
        # TODO: layer 'Mean Pool' (globalAvgPool1d) is not yet supported by the exporter; passing through unchanged
        linear_ser_fc = self.linear_1(x)
        concatenate__final = torch.cat([transformer_block_trf, linear_ser_fc], dim=-1)
        linear_mlp1 = self.linear_2(concatenate__final)
        prelu_act1 = self.prelu_1(linear_mlp1)
        linear_mlp2 = self.linear_3(prelu_act1)
        prelu_act2 = self.prelu_2(linear_mlp2)
```

## Machine access

- Every architecture: https://neurarch.com/a/index.json
- Verify a graph of your own: `POST https://www.neurarch.com/api/v1/check` (see https://neurarch.com/developer.html)
- MCP server, so an agent edits the graph with the checks in the loop: https://neurarch.com/docs/mcp.md