智能风控原理、算法与工程实践 (梅子行)（Z-Library）

【One-Line Pitch】 A practitioner-oriented guide to building machine-learning risk-control systems for consumer credit, showing how to turn messy lending data into fraud and default predictions that survive real deployment. Best for analysts, modelers, and engineers who already know basic credit scorecards and want the advanced, code-heavy playbook. 【Book Arc】 - **Opening (~0%–10%)**: Frames why finance leans on Python and traces risk management from pure human judgment to "human + computer" to today's intelligent stage; introduces the four defining traits of risk-control ML (rare negatives, high explainability demands, diverse models, rich data) and the end-to-end industrial modeling workflow. - **Early (~10%–30%)**: Covers credit fundamentals, label definition, sampling, observation/performance windows, and the develop/validation/out-of-time split; then builds rule engines by combining feature engineering with CART regression trees to auto-mine segmentation rules. - **Early–Middle (~30%–45%)**: Moves into ensemble scorecards with XGBoost, contrasting complex-feature-plus-simple-algorithm against simple-feature-plus-complex-algorithm, and details credit feature families (personal, financial, multi-lending, consumption, platform history, app tracking, external credit bureau). - **Middle (~45%–60%)**: Dives into feature engineering mechanics — time-window aggregations, trend/volatility features — and categorical encoding via one-hot versus WOE, including automated WOE binning that maximizes negative-rate separation. - **Late (~60%–85%)**: Addresses transfer learning for project cold start, survivor bias and imbalanced learning through unsupervised and semi-supervised methods, and unsupervised anomaly detection for data cleaning and individual fraud detection. - **Ending (~85%–100%)**: Shares proven model-optimization and model-fusion strategies, then closes with knowledge-graph and complex-network basics plus network representation learning, where community detection supports gang-fraud detection and credit-model improvement. 【Key Takeaways】 - **Risk-control ML is defined by its constraints, not its algorithms** (Opening): rare negatives, explainability pressure, and business-specific models shape every design choice, so generic ML recipes rarely transfer directly. - **Label definition is a modeling decision, not a given** (Early): using a 15-day overdue threshold and dropping ambiguous "gray samples" makes the binary distribution cleaner and the model easier to learn. - **Sampling must be paired with weighting** (Early): undersampling positives without re-adding weights distorts validation metrics and expected loss, so weights must flow through evaluation. - **Rules and models solve different problems** (Early): rule mining via decision trees delivers fast, interpretable segmentation, while ML models suit precision-critical cases — mature teams use both. - **Feature engineering dominates outcomes** (Middle): time-window aggregation, trend, and volatility features around repayment ability and willingness matter more than tuning, and WOE encoding beats one-hot for short categorical variables. - **Imbalance and cold start are first-class problems** (Late): transfer learning, semi-supervised, and unsupervised anomaly methods are the practical answers when labels are scarce or a new product has no history. - **Fraud detection is not binary** (Late): because fraud tactics vary, it behaves more like multi-class or anomaly detection, and graph/community methods are needed for organized fraud rings. - **Model fusion and optimization yield incremental, not transformative, gains** (Ending): better variables and model selection beat parameter tuning, and complex models are often best used as feature generators feeding simpler scorecards. 【Reading Tips】 - Deep-read Chapters 1–2 if you are new to credit modeling; the workflow, label, and sampling logic underpins everything later. - Skim the long code blocks on first pass — extract the design pattern (aggregation function, binning rule) rather than memorizing syntax. - Treat the WOE binning and decision-tree rule-mining sections as templates you can adapt to your own data. - For the later chapters on transfer learning, anomaly detection, and graph methods, focus on when to apply each technique rather than full implementation detail. - Keep the business framing in view: every algorithm here exists to serve a lending decision, so ask "what decision does this improve?" 【Coverage Limits】 This guide is synthesized from stratified excerpts covering roughly the first half of the book in detail, with later chapters (transfer learning, imbalance handling, anomaly detection, model fusion, knowledge graphs) represented mainly by chapter-level summaries; specific case results and full code in those later chapters are not covered.