Harshvardhan226510/Core-Machine-Learning-Toolkit

# 核心 Machine Learning 工具包 欢迎使用核心 Machine Learning 工具包。本仓库是一个工程展示项目，包含了基础的 machine learning 实现、稳健的 preprocessing 自动化以及优化的 model pipeline。它涵盖了生产级工作流，涉及高维 classification、不平衡的金融 data engineering、空间 feature 生成以及实时交互式用户服务。 ## 📁 仓库目录结构 本仓库明确划分为四个核心预测领域进行模块化： ``` ├── california-housing/ │ └── housing_regressor.ipynb # Spatial transformation & XGBoost regression pipeline ├── crop-prediction/ │ └── crop_recommendation.ipynb # IoT sensor optimization & stratified ensemble classifier ├── fraud-detection/ │ └── financial_fraud_detector.ipynb # Imbalanced transaction modeling & Gradio web application ├── hotel-cancellation/ │ └── cancellation_predictor.ipynb # Complex feature encoding, outlier management & classification └── README.md # Central portfolio documentation matrix ``` ## 🌾 1. 智能作物推荐系统 ``` An industrial-grade agro-analytics pipeline designed to optimize agricultural yields by matching soil chemistry profiles with macro-climatic conditions. Exploratory Data Analysis: Profiled soil Nitrogen (N), Phosphorus (P), and Potassium (K) distributions using multi-variate box plots alongside soil pH bounds to validate physical IoT sensor utility. Pipeline Engineering: Wrapped automated feature scaling (StandardScaler) and multi-class tracking configurations inside an encapsulated Scikit-Learn pipeline. Core Intelligence: Leveraged an XGBClassifier to process 2,200 rows across 22 distinct crop varieties, scoring a 98.86% base accuracy validated by a highly stable 5-Fold cross-validation benchmark (99.09% to 99.77%). ``` ## 🏠 2. 加州房价 Regressor ``` An advanced geographic and spatial regression ecosystem engineered to evaluate multi-variate regional real estate valuation targets. Feature Engineering & Imputation: Addressed missing data points using a localized median SimpleImputer. Utilized a custom vectorized FunctionTransformer to synthesize real-time structural ratio indicators: Rooms per Household = Total Rooms / Households Bedrooms per Household = Total Bedrooms / Households Population per Household = Population / Households Encoding Matrix: Implemented an explicit OrdinalEncoder to transform string-based coastline proximity values (ocean_proximity) based on real-world categorical indexing boundaries. Hyperparameter Optimization: Conducted high-throughput hyperparameter optimization utilizing GridSearchCV over an estimators, depth, and learning rate grid, matching a strong operational R² accuracy score of 0.85. ``` ## 💳 3. 信用卡欺诈检测系统 ``` An enterprise financial intelligence system built to analyze multi-million row transaction logs and flag fraudulent behavior inside highly imbalanced environments. Class Imbalance Mitigation: Handled extreme target imbalance (over 6.3M clean transactions vs. just 5,742 fraud data records) by calculating the strict imbalance factor matrix (770) and mapping it straight into XGBoost's structural weight constraints (scale_pos_weight=500). Performance Matrix: Optimized model parameters heavily focusing on defensive recall thresholds, resulting in an exceptional 98% raw fraud recall rate to guarantee comprehensive risk interception. Gradio Application Serving: Serialized the complete extraction pipeline wrapper (Fraud_prediction_model.pkl) and bound it to an interactive web UI layout using the Gradio framework for live transaction inference tracking. ``` ## 🏨 4. 酒店预订取消预测器 ``` A behavioral analysis model engineered to predict operational room churn and reservation cancellations across 119,390 transaction rows. Data Cleaning & Outlier Squashing: Stripped systemic leakage factors and uninformative columns. Addressed high-variance financial skewing by utilizing median bounding masks on the Average Daily Rate (adr) column to neutralize erratic anomalies (< 0 or > 900). Mixed-Type Column Transformer: Constructed a comprehensive data transformation network routing features down three explicit avenues: Numerical Avenue: Median imputation paired with standard normalization scaling. One-Hot Avenue: Most-frequent imputation string strategy paired with highly sparse structural encoding matrices. Ordinal Avenue: Explicit layout maps translating booking room types and refundable token properties. Ensemble Optimization: Fine-tuned an intensive XGBClassifier loop via GridSearchCV over parameters to achieve an optimal production testing score of 88.02% accuracy. ```

标签：Apex, NoSQL, Scikit-learn, XGBoost, 数据科学, 机器学习, 模型 pipeline, 资源验证, 逆向工具, 预测分析