本项目是 IS6400 课程的小组项目，专注于商业数据分析在金融风控领域的应用。我们构建了一个综合的机器学习框架，主要解决贷款违约预测这一关键的金融风控问题。
主要项目：
- 贷款违约预测 - 通过多种机器学习模型识别潜在的违约客户（最终成品）
前期探索：
- 汽车价格预测 - 前期尝试的项目，后期已放弃，相关代码保留作为学习参考
- 风险控制：提高银行识别违约客户的能力，降低信贷损失
- 模型比较：评估不同机器学习算法在金融风控场景中的表现
- 业务应用：为银行信贷审批提供可靠的决策支持工具
- 贷款数据集：来自 Kaggle Loan Approval Classification Data，包含客户基本信息、信贷历史、财务状况等特征（主要数据集）
- 汽车价格数据集：来自 Kaggle Car Price Prediction Challenge（前期探索，已放弃）
- 编程语言：Python 3.8+
- 机器学习：Scikit-learn, CatBoost, XGBoost
- 深度学习：TensorFlow/Keras, BERT
- 数据处理：Pandas, NumPy
- 可视化：Matplotlib, Seaborn
- 开发环境：Jupyter Notebook
- 逻辑回归 (Logistic Regression)
- K近邻算法 (KNN)
- 决策树 (Decision Tree)
- 随机森林 (Random Forest)
- CatBoost - 梯度提升算法
- Stacking - 多模型堆叠集成
- Voting - 投票集成
- 多层感知机 (MLP/Neural Network)
- BERT - 基于Transformer的语言模型
- 聚类分析 (Clustering Analysis)
- 特征选择 (向前选择/向后消除)
is6400-business-data-analytics/
│
├── data/ # 数据文件
│ ├── loan_data.csv # 原始贷款数据
│ ├── cleaned_loan_data.csv # 清洗后的贷款数据
│ ├── further_cleaned_dataset.csv # 进一步清洗的数据
│ └── car_price_prediction.csv # 汽车价格数据（前期探索）
│
├── data_cleaning/ # 数据清洗
│ ├── loan_data_cleaning.ipynb # 贷款数据清洗
│ └── further_data_cleaning.ipynb # 深度数据清洗
│
├── loan_data_analytics/ # 贷款违约预测模型（核心项目）
│ ├── logistic and cluster.ipynb # 逻辑回归与聚类
│ ├── knn.ipynb # K近邻算法
│ ├── decision_tree.ipynb # 决策树
│ ├── random_forest.ipynb # 随机森林
│ ├── catboost.ipynb # CatBoost模型
│ ├── bert.ipynb # BERT模型
│ ├── stacking.ipynb # Stacking集成
│ ├── boost_voting.ipynb # 提升投票算法
│ ├── stacking_results.md # Stacking结果分析
│ ├── 向前选择和向后消除.md # 特征选择方法
│ └── image/ # 结果图像
│
├── data_analytics/ # 早期汽车价格预测探索（已放弃）
│ ├── multi_linear_reg.ipynb # 多元线性回归
│ ├── mlp.ipynb # 多层感知机
│ ├── the_boss_mlp.ipynb # 优化版MLP
│ ├── catboost.ipynb # CatBoost回归
│ ├── best_model.h5 # 神经网络模型
│ └── best_catboost_model.cbm # CatBoost模型
│
├── README.md # 项目说明文档
└── LICENSE # 开源许可证
| 模型 | 准确率 | 召回率(违约) | 精确率(违约) | F1分数 | False Negatives |
|---|---|---|---|---|---|
| CatBoost | 94% | 80% | 90% | 0.85 | 394 |
| Random Forest | 93% | 76% | 90% | 0.83 | 477 |
| Decision Tree | 90% | 78% | 77% | 0.77 | 450 |
| KNN | 90% | 72% | 80% | 0.76 | 550 |
| BERT | 93% | 80% | 90% | 0.85 | 394 |
CatBoost模型表现最佳：
- ✅ 最高的整体准确率 (94%)
- ✅ 最好的违约客户识别能力 (80% 召回率)
- ✅ 最少的风险遗漏 (394个False Negatives)
- ✅ 高精确率 (90%) 减少误伤优质客户
- 风险控制提升：相比基准模型，CatBoost减少了156个潜在坏账 (550→394)
- 机会成本优化：高精确率减少了对优质客户的误判
- 决策支持：为银行信贷审批提供可靠的风险评估工具
- 多模型比较：系统性比较了从传统统计方法到深度学习的多种算法
- 集成学习：通过Stacking技术融合多个模型的优势
- 特征工程：实施了向前选择和向后消除等特征选择方法
- 业务导向：分析结果紧密结合实际业务场景和决策需求
本项目由IS6400课程小组协作完成，涵盖数据科学、机器学习和商业分析等多个领域。
本项目采用 MIT 许可证 - 详见 LICENSE 文件
This project is a group assignment for the IS6400 course, focusing on business data analytics applications in financial risk control. We have built a comprehensive machine learning framework to address the critical business problem of loan default prediction.
Main Project:
- Loan Default Prediction - Identifying potential defaulting customers through various machine learning models (Final Product)
Early Exploration:
- Car Price Prediction - An early exploration project that was later abandoned, with related code retained for learning reference
- Risk Control: Enhance banks' ability to identify defaulting customers and reduce credit losses
- Model Comparison: Evaluate the performance of different machine learning algorithms in financial risk control scenarios
- Business Application: Provide reliable decision support tools for bank credit approval
- Loan Dataset: From Kaggle Loan Approval Classification Data, contains customer demographics, credit history, financial status, and other features (Primary Dataset)
- Car Price Dataset: From Kaggle Car Price Prediction Challenge (Early exploration, abandoned)
- Programming Language: Python 3.8+
- Machine Learning: Scikit-learn, CatBoost, XGBoost
- Deep Learning: TensorFlow/Keras, BERT
- Data Processing: Pandas, NumPy
- Visualization: Matplotlib, Seaborn
- Development Environment: Jupyter Notebook
- Logistic Regression
- K-Nearest Neighbors (KNN)
- Decision Tree
- Random Forest
- CatBoost - Gradient Boosting Algorithm
- Stacking - Multi-model Stacking Ensemble
- Voting - Voting Ensemble
- Multi-Layer Perceptron (MLP/Neural Network)
- BERT - Transformer-based Language Model
- Clustering Analysis
- Feature Selection (Forward Selection/Backward Elimination)
| Model | Accuracy | Recall (Default) | Precision (Default) | F1-Score | False Negatives |
|---|---|---|---|---|---|
| CatBoost | 94% | 80% | 90% | 0.85 | 394 |
| Random Forest | 93% | 76% | 90% | 0.83 | 477 |
| Decision Tree | 90% | 78% | 77% | 0.77 | 450 |
| KNN | 90% | 72% | 80% | 0.76 | 550 |
| BERT | 93% | 80% | 90% | 0.85 | 394 |
CatBoost Model Performs Best:
- ✅ Highest overall accuracy (94%)
- ✅ Best default customer identification capability (80% recall)
- ✅ Fewest risk omissions (394 False Negatives)
- ✅ High precision (90%) reduces misclassification of quality customers
- Risk Control Improvement: CatBoost reduced 156 potential bad loans compared to baseline (550→394)
- Opportunity Cost Optimization: High precision reduces misclassification of quality customers
- Decision Support: Provides reliable risk assessment tools for bank credit approval
- Multi-model Comparison: Systematic comparison of algorithms from traditional statistics to deep learning
- Ensemble Learning: Leveraging advantages of multiple models through Stacking techniques
- Feature Engineering: Implementation of forward selection and backward elimination methods
- Business-Oriented: Analysis results closely integrated with actual business scenarios and decision needs
This project is collaboratively completed by the IS6400 course group, covering multiple fields including data science, machine learning, and business analytics.
This project is licensed under the MIT License - see the LICENSE file for details
如有问题或建议，欢迎提交Issue或Pull Request。 For questions or suggestions, feel free to submit an Issue or Pull Request.
⭐ 如果这个项目对您有帮助，请给我们一个Star！ ⭐ If this project helps you, please give us a Star!