Ecommerce-Shipping-Predictive-Modeling

Data source : https://www.kaggle.com/prachi13/customer-analytics

Project Overview

• Seek insight from the dataset with Exploratory Data Analysis
• Performed data cleansing, data processing, data engineering to prepare data before modeling
• Built a model to predict whether the shipping deliveries will be received late or on time by the customers
• Developt recommendations & benefit analysis based on insights and model prediction

Background & Problems

An international e-commerce company that sells electronic products want to discover key insights from their customer database. Currently, most of the shipping deliveries are late.

Exploratory Data Analysis

Variable	Type	Definition	Example
ID	Nominal	Customer ID Number	10, 15, 10995, 10996
Warehouse_block	Nominal	Warehouse to Store the Product	A, B, C, D, F
Mode_of_Shipment	Nominal	Mode of Product Shipping	Flight, Road, Ship
Customer_care_calls	Discrete	Number of Calls Made	1, 2, 5, 6
Customer_rating	Ordinal	Company Rating by Customers	5: Best - 4: Better - 3: Neutral - 2: Bad - 1: Worst
Cost_of_the_Product	Discrete	Cost of Product in US Dollars	177, 216, 236, 182
Prior_purchases	Discrete	Number of Prior Purchase	3, 2, 6
Product_importance	Ordinal	Product Importance Parameter	Low, Medium, High
Gender	Nominal	Customer Gender	Male, Female
Discount_offered	Discrete	Product Discount in US Dollars	65, 10, 16
Weight_in_gms	Continous	Product Weight in grams	4953, 5676, 2171
Reached.on.Time_Y.N	Nominal	Target Variable, 1: NOT reached on time - 0: REACHED on time	1, 0

1. 59.7% of e-commerce shipping deliveries are late received by the customers (6.563 of 10.999 customers).

2. Ship & Warehouse F has the highest frequency of delivery. But it looks almost the same based on the percentage. There's an assumtion that the late is influenced by other factors.

3. Every product that gets a discount above 10 is confirmed Late. There is an assumption that this happens in specific months, but needs further checking.

4. Shipping delivery is confirmed late when the product weight is between 2-4 kg.

Data Processing

• Check missing & duplicate values
• Remove outliers for Purchase feature by IQR limit & replace outliers for Discount feature with IQR limit
• Ordinal encoding for Importance column & feature encoding the rest of categorical columns
• Select best features for modeling
• Normalize all selected features

Modeling

• Split features & target
• Split data into data train & data test
• Train model with 5 different algorithm such as Logistic Regression, Random Forest, XGBoost & AdaBoost
• Evaluate model with Recall, Average Precision (AP) & ROC Score
• Validate model with train, test accuracy & AP train, AP test as well
• Hyperparameter tuning
• Select the best model

Model Evaluation

Score	Random Forest	Logistic Regression	AdaBoost	XGBoost
Recall	0.57	0.73	0.60	0.68
AP	0.74	0.67	0.70	0.68
ROC	0.70	0.62	0.68	0.66
AP Train	0.74	0.67	0.72	0.93
AP Test	0.74	0.67	0.70	0.68

Based on the model evaluation, Random Forest algorithm has the best result.

Recommendations

Short terms

• Add estimatedarrival time to assure the package arrived on time
• Give credit points as a compensations to retain customer loyalty

Long terms

• Add more features to give more specific & accurate insights
• Perform operational audit based on the insights