Fix: Restore YAML frontmatter for Gradio SDK configuration

README.md

@@ -1,7 +1,17 @@
 ---
+sdk: gradio
+title: Insurance App Behavior
+emoji: 🏥
+colorFrom: purple
+colorTo: gray
+sdk_version: 6.14.0
+python_version: '3.13'
+app_file: app.py
+pinned: false
 tags:
 - ml-intern
 ---
+
 # 🏥 保险APP 用户行为分析模型训练平台 v3.0
 
 基于最新研究论文构建的**工业级保险用户行为分析平台**，支持**7大功能模块**：演示模式、CSV上传、产品推荐、异常检测、模型管理、生存分析、帮助文档。
@@ -39,20 +49,6 @@ tags:
 - **输出**: 购买概率 + 注意力权重可视化
 - ** insight**: 用户兴趣表示随候选产品动态变化
 
-```
-用户历史: [event_1, product_1], [event_2, product_2], ...
-       ↓ Embedding
-事件嵌入(D/2) + 产品嵌入(D/2) → 行为嵌入(D)
-       ↓
-候选产品 Embedding ───┐
-                       ↓
-         [c, b, c-b, c*b] → Attention MLP → 权重 α
-                       ↓
-            加权求和 → 兴趣向量(D)
-                       ↓
-   [用户, 兴趣, 候选, 交互, 统计特征] → MLP → 购买概率
-```
-
 ### 3. 异常行为检测 (TabBERT)
 **方法**: TabularBERT + Focal Loss
 - **核心**: 层次化Transformer + 不平衡数据处理
@@ -125,14 +121,6 @@ pip install -r requirements.txt
 python app.py
 ```
 
-### Docker 运行
-
-```bash
-docker run -p 7860:7860 --platform=linux/amd64 \
-    -e HF_TOKEN="your_token" \
-    registry.hf.space/stephanwu-insurance-app-behavior:latest
-```
-
 ---
 
 ## ⚠️ 不平衡数据处理
@@ -161,73 +149,6 @@ Stephanwu/insurance-app-behavior/
 
 ---
 
-## 🧠 模型架构详解
-
-### DIN (Deep Interest Network)
-```python
-# LocalActivationUnit 核心
-candidate_emb = embed(candidate_product)          # (B, D)
-behavior_emb = embed(events) + embed(products)    # (B, L, D)
-
-# 4路交互特征
-interaction = concat([
-    candidate_emb,           # 候选产品
-    behavior_emb,            # 历史行为
-    candidate - behavior,     # 差异
-    candidate * behavior,    # 点积
-])  # (B, L, 4D)
-
-# 注意力权重
-attention_weights = MLP(interaction)  # (B, L)
-attention_weights = softmax(attention_weights)
-
-# 加权兴趣
-interest = sum(behavior_emb * attention_weights)  # (B, D)
-
-# 预测
-logits = MLP(concat([user, interest, candidate, interaction, stats]))
-```
-
-### TabBERT (简化版)
-```python
-# 层次化Transformer
-input_features = [claim_amount, claim_type, days_since_policy, ...]
-          ↓
-    Linear Projection: d_model (128)
-          ↓
-    ┌────────────────────────┐
-    │  Transformer × 4       │  # 模拟 Field + Sequence level
-    │  LayerNorm + Dropout    │
-    └────────────────────────┘
-          ↓
-    Global Average Pooling
-          ↓
-    MLP: 128 → 256 → 64 → 1
-          ↓
-    Focal Loss (解决1:4不平衡)
-```
-
-### DeepSurv (Neural Cox-PH)
-```python
-# Cox partial likelihood loss
-def cox_ph_loss(pred, time, event):
-    # Sort by time descending
-    pred_sorted = pred[argsort(time, descending=True)]
-    event_sorted = event[argsort(time, descending=True)]
-    
-    # logcumsumexp for numerical stability
-    log_cumsum_h = logcumsumexp(pred_sorted)
-    
-    # Only event samples contribute
-    loss = -sum(event * (pred - log_cumsum_h)) / sum(event)
-    return loss
-
-# Survival probability
-S(t | x) = exp(-H_0(t) * exp(pred(x)))
-```
-
----
-
 ## 📚 参考文献
 
 | 论文 | 应用 | arXiv | 会议 |