How to Leverage Big Data to Deliver Smart Logistics
Download as PPTX, PDF5 likes505 views
The document discusses leveraging big data and AI solutions in the logistics industry, focusing on demand prediction, inventory management, and delivery optimization. It highlights the Tianchi big data contest platform, which connects data talents and provides practical solutions for logistics challenges. Finally, it emphasizes the importance of meteorology in optimizing delivery routes for air transportation.
How to Leverage Big Data to Deliver Smart Logistics
- 1. Presenter: Rui Xiong January 2018 How to Leverage Big Data to Deliver Smart Logistics
- 2. 1 2 3 4 What You Will Learn Background on the Logistics Industry Future Challenge: Helping Balloons Navigate the Weather AI Solutions in Logistics & Delivery Optimization Tianchi Big Data Contest Platform www.alibabacloud.com 2 / 32
- 3. www.alibabacloud.com 01 Background on the Logistics Industry 3 / 32
- 4. www.alibabacloud.com 18% 8.7% 8.5% Industry Background Logistics Cost / GDP ￥13 Trillion/ USD $2 Trillion • Transportation: 54.1% • Inventory Management: 33.2% • Administration: 12.7% Reference: Yu, Chunge. "The Analysis of the China National Logistics Costs Structure." Management & Engineering 21: 77. 4 / 32
- 5. Intelligent logistics Cost Efficiency Demand Delivery Inventory Predictable Controllable Reachable Objectives of Intelligent Logistics www.alibabacloud.com 5 / 32
- 6. www.alibabacloud.com 02 AI Solutions in Logistics & Delivery Optimization 6 / 32
- 7. Intelligent Logistics Demand Prediction Inventory Management Delivery Optimization • Daily, weekly, monthly prediction • Different scenarios • What, when and how much to replenish • Improve inventory turnover rate • Vehicle-cargo matching • Route planning • Distribution network layout design • Warehouse coverage of facilities • Inventory setting for each warehouse Overall Supply Chain Optimization Overview of Intelligent Logistics www.alibabacloud.com 7 / 32
- 8. High cost • large numbers of couriers Low efficiency • random dispatch, backlog of orders Intelligent dispatch Route planning Integrated management 9 million order processing capability 500 millisecond response time 29 minute average delivery time Hundred of millions RMB cost reduction Case Study: Ele.me Problem Solution Result www.alibabacloud.com 8 / 32
- 10. Route Calculation Vehicle, Driver, Cargo, Traffic, Weather, Time, etc. User-loss Analysis Matching Module Basic Feature Analysis Driver Evaluation Order Evaluation Data Order PredictionComplex Feature Analysis Vehicle-Cargo Matching Historical Recommendation Engine Real Time+ www.alibabacloud.com 10 / 32
- 11. Order Prediction www.alibabacloud.com under-fitover-fit • Time Series Model: AR Model • Machine Learning Model: Linear, Non-Linear Models 11 / 32
- 12. Demand Prediction www.alibabacloud.com Classification Regression Model Rules Commodities Set Merge Commodities With Abnormal Sales Commodities With Normal Sales 12 / 32
- 13. www.alibabacloud.com • Root Mean Square Error (RMSE) decreases by 20-25% compared to regression • Robustness and strong generalization Demand Prediction 13 / 32
- 14. Route Planning www.alibabacloud.com • Shortest Path Problem a) Dijkstra's Algorithm b) A* Algorithm • Vehicle Routing Problem (VRP) a) Genetic Algorithm b) Adaptive Large Neighborhood Search Algorithm http://home.ku.edu.tr/~daksen/ISOLDE2014-ALNS-for-SPIRP.pdf 5 2 7 10 18 22 66 2010 20 16 4 6 8 24 20 6 9 14 / 32
- 15. User’s Transportation Management System Facility, Order, Vehicle, Historical Data, etc. Execution Data Cleaning And Mining Vehicle-routing Algorithm User’s Data External data RDS Alibaba Cloud Relational Database Service Route Optimization Engine MaxCompute Alibaba Cloud’s Big data computing ① ② ③ Human Adjustment Final Delivery Plan AMAP YUNDI Delivery Plan www.alibabacloud.com Route Planning: Framework 15 / 32
- 16. www.alibabacloud.com Route Optimization Gantt Chart for Vehicle Arrangement 3D Cargo Layout Manual Adjustment & Result Confirmation 1 3 2 4 Optimized Route Plan Route Planning: Core Functions 16 / 32
- 17. www.alibabacloud.com 03 TIANCHI Big Data Contest Platform 17 / 32
- 18. www.alibabacloud.com https://tianchi.aliyun.com/?_lang=en_US A global contest platform focused on big data A community of data talents: 110,000+ developers 2,800+ academic institutes & business groups 75 countries and regions worldwide Provides professional solutions What is TIANCHI 18 / 32
- 19. www.alibabacloud.com Crowd- sourcing Data Sets Forum ContestsA world leaderboard of data talents Practical solutions from collective input Data sets and computing resources A Learn-Explore-Share one-stop community What TIANCHI Offers 19 / 32
- 20. www.alibabacloud.com DATA MINING KDD CUP 2017 Highway Tollgates Traffic Flow Prediction CIKM AnalytiCup 2017 Short-term Quantitative Precipitation Forecasting OPERATION RESEARCH THE MET OFFICE 2017-2018 Future Challenge—Helping Balloons Navigate the Weather INFORMS 2016 The Last Mile Rush ARTIFICIAL INTELLIGENCE IJCAI 2017 Customer Flow Forecasts on Koubei.com IJCAI 2016 Brick-and-Mortar Store Recommendation with Budget Constraints IJCAI 2015 Repeat Buyers Prediction after Sales Promotion TIANCHI International Contests 20 / 32
- 21. www.alibabacloud.com • Estimate the average travel time from designated intersections to tollgates • Predict average tollgate traffic volume Task • The road network topology in target areas • Vehicle trajectories • Historical traffic volume at tollgates • Weather data (Nasdaq AMAP and the government) KDD CUP 2017 Data 21 / 32
- 22. www.alibabacloud.com • Estimate the average travel time from designated intersections to tollgates • Predict average tollgate traffic volume • The road network topology in target areas • Vehicle trajectories • Historical traffic volume at tollgates • Weather data ( Nasdaq AMAP and the government) KDD CUP 2017 Task 1: To estimate the average time from designated intersections to tollgates Source: http://www.kdd.org/kdd2017/announcements/view/announcing-kdd-cup-2017-highway-tollgates-traffic-flow-prediction 22 / 32
- 23. www.alibabacloud.com Source: http://www.kdd.org/kdd2017/files/Task1_1stPlace.pdf Solution From Team Convolution 23 / 32 Feature Engineering • More than 100 features extracted • Feature selection based on CV experiments and online feedback • Multiple feature combinations used for ensemble, solving feature conflict
- 24. www.alibabacloud.com Source: http://www.kdd.org/kdd2017/files/Task1_1stPlace.pdf 24 / 32 Tree Based Model XGBOOST • Level-wise growth strategy • Stable LightGBM • Leaf-wise growth strategy • Good algorithm for category features • Fast Model Tuning • Feature selection based on CV experiments and online feedback • Multiple feature combinations used for ensemble, solving feature conflict Sample Weight • Use 1/label as sample weight, make the little values learn better Solution From Team Convolution
- 25. www.alibabacloud.com Source: http://www.kdd.org/kdd2017/files/Task1_1stPlace.pdf 25 / 32 One Hot Embedding • Using embedding layer for basic feature • Similar locations’ embedding vector is close Early Interaction vs Later Interaction • The balance of learning between embedding and statistics MLP vs RNN • Multiple Layer Perception: Powerful Expression Ability, Learn Feature Interaction • Recurrent Neural Network: Good to model sequence relationships but unstable for generalization Solution From Team Convolution
- 26. www.alibabacloud.com Source: http://www.kdd.org/kdd2017/files/Task1_1stPlace.pdf 26 / 32 Model Level Ensemble • Different ML models, including xgb, lightbm, dnn, rnn • Loss function changes, weighted norm-2 and fair loss to approximate norm-1 loss • Label transform, such as customized log transform and sample weight transform • Model parameter, such as nn structure, gbdt parameters Results • Chose 13 base models based on multiple weighted cv and leaderboard feedback • Achived 0.1748 mean absolute percentage error (MAPE) on leaderboard Solution From Team Convolution
- 27. www.alibabacloud.com • The Olympic Games in the field of data mining • 3582 teams participated, the most of all previous KDD Cups KDD CUP 2017 27 / 32
- 28. www.alibabacloud.com 04 Future Challenge—Helping Balloons Navigate the Weather 28 / 32
- 29. www.alibabacloud.com • Business demands for delivery route optimization • Importance of meteorology to air transports • Future Prospect
- 30. www.alibabacloud.com Q & A 30 / 32
- 31. www.alibabacloud.com • www.alibabacloud.com • www.tianchi.aliyun.com/?_lang=en_US • www.alibabacloud.com/campaign/bigdatacontest ruixiong.srx@alibaba-inc.com Further Resources 31 / 32
Editor's Notes
- #6: Inventory management Replenishment decision-making support Improve inventory turnover rate Reasonable warehouse layout/distribution Delivery planning and control (Trunk/Branch-line transportation, city distribution, last-mile delivery ) Synthesis of data from warehouse, cargo, route, vehicle, and order Precise vehicle-cargo matching Efficient route planning Flexible real-time scheduling Delivery solution for various scenes of trunk/branch-line transportation, city distribution, last-mile delivery Smooth connection between various sectors or warehouses Demand prediction Accurate demand prediction From human-seeking-good to good-seeking-human
- #9: 问题： 成本高：订单量大，派单人员多 效果差：人工派单靠经验，随机性强 效率低：订单积压严重，时效性差 方案： 智能派单: 基于画像分析，标签信息 路径规划: 基于订单、实时地图信息 整合管理: 整合设备数据，监控配送 结果： 900万 高峰订单处理能力 500毫秒 响应时间 29分钟 平均配送时间 数亿元 人力成本降低