E-commerce Logistics Delay Prediction: A Practical Machine Learning Case Study Based on CRISP-DM

This project is the final data science project of Group 4 in KKI ITDS. Using the CRISP-DM methodology, it builds a logistics delay prediction model based on 10,999 order records. After comparing three algorithms—Decision Tree, Random Forest, and KNN—it recommends the Random Forest model and identifies discount offers and product weight as the most critical factors for delay prediction. The project source is GitHub (link: https://github.com/group4-kki-itds/intro-to-data-science-final-project-group-4-kki-2026), published on June 5, 2026.

The booming development of e-commerce has led to exponential growth in logistics scale. Logistics delays affect customer satisfaction and increase enterprise operating costs (such as re-delivery, complaint handling, order cancellation, etc.). The goal of this project is to build an order delay prediction model, which helps enterprises take proactive measures (priority processing, changing logistics channels, early communication, etc.) to reduce the delay rate by identifying high-risk orders in advance.

The project strictly follows the CRISP-DM (Cross-Industry Standard Process for Data Mining) framework, which includes six phases:

1. Business Understanding: Clarify the core goal of predicting order delays and the success criteria;
2. Data Understanding: Use a dataset containing 10,999 order records, covering dimensions such as order attributes, product information, and logistics details;
3. Data Preparation: Perform data cleaning, feature engineering, missing value handling, and categorical variable encoding;
4. Modeling: Compare three classic machine learning algorithms;
5. Evaluation: Comprehensive evaluation of model performance using multiple metrics;
6. Deployment Recommendations: Provide a plan for integrating the model into business processes.

The project compares three classification algorithms:

• Decision Tree: Strong interpretability, but prone to overfitting and limited generalization ability;
• KNN: Instance-based learning, sensitive to feature scaling, and performance may decline in high-dimensional data;
• Random Forest: Integrates multiple decision trees, has higher stability and accuracy, is robust to noise and outliers, and can achieve good results without complex parameter tuning. Finally, the Random Forest model is recommended, which aligns with industry practices for handling tabular data classification tasks.

Feature importance analysis reveals two key predictive factors:

1. Discount Offers: The most important factor. The surge in orders during promotions leads to insufficient logistics capacity, or discounted products come from different warehouses/use different logistics strategies;
2. Product Weight: The second most important factor. Heavier products require special logistics arrangements, longer processing times, or are subject to transportation restrictions (e.g., cannot be shipped by air). These findings provide actionable insights for enterprises: increase logistics resources during promotion periods and handle heavy cargo orders differently.

Based on the model results, the project puts forward the following recommendations:

1. Early Warning Mechanism: Conduct delay risk scoring when orders enter the system, and trigger special attention for high-risk orders;
2. Resource Allocation: Predict high-delay periods (e.g., promotion periods) and increase logistics resources in advance;
3. Customer Communication: Proactively communicate the expected delivery time for high-risk orders to manage customer expectations;
4. Process Optimization: Design differentiated processing procedures for heavy cargo and discount orders.