This project aims to study the shipping status from an e-commerce company, aiming to identify and understand late deliveries to improve customer experience. The dataset is obtained from Kaggle, and the analysis notebook is uploaded to Google Colab
- Data Wrangling with pandas
- EDA with seaborn
- Feature Selection and Scaling
- Sample Balancing with RandomOverSampler
- Classification with Machine Learning algorithms, Ensemble Methods, and Artificial Neural Network
The dataset contains 12 columns, with a total of 10999 rows. The target variable is Reached.on.Time_Y.N.
| Variable Name | Description | Data Type |
|---|---|---|
| ID | Unique ID for each shipment | int64 |
| Warehouse_block | Name of warehouse block (A,B,C,D, F) | object |
| Mode_of_Shipment | Mode of shipment (Flight, Road, Ship) | object |
| Customer_care_calls | Number of calls from customer | int64 |
| Customer_rating | Customer's shipment rating (1-5) | int64 |
| Cost_of_the_Product | Product's prices | int64 |
| Prior_purchases | Number of previous purchases | int64 |
| Product_importance | Degree of importance (low, medium, high) | object |
| Gender | Customer's gender | object |
| Discount_offered | Discount offered on that specific product | int64 |
| Weight_in_gms | Product's weight in grams | int64 |
| Reached.on.Time_Y.N | Whether the shipment arrived as scheduled (0,1) | int64 |
| - Shipping_Ecommerce
| -- data Contains the raw dataset
| --- customer_data.csv
| -- E-commerce Shipping Analysis - Project Report.pdf Contains details about the analysis with findings
| -- E_commerce_shipping_analysis.ipynb Contains the source codes
| -- LICENSE MIT License
| -- README.md Project Overview
As we carried out our feature selection step, it was apparent that only 2 variables: Discount_offered and Weight_in_gms seem to have determining impact on shipment's status. Due to the lack of additional domain knowledge, it is challenging to explain the reasons behind this issue.
However, we can infer that, although the classification task with two independent variables is simpler, it still presents a probabilistic problem that benefits from using Machine Learning. While the simple thresholds of these two variables can be used to flag any potentially delayed order, the fact that we are providing an accurate model that can analyze each case by their specific value and render probabilistic outputs makes this process more efficient and precise. Moreover, our report explores the instrinsic relationship between these variables and other features, presenting the argument that they also carry implicit information regarding Product_importance.
A final consideration to be made here is that additional inputs to the model would be beneficial, particularly by aggregating data related to the shipment contracts and specifications with the supplier for each order. We believe that this would aid the models in interpreting the impacts of the actual shipping to delivery times. We hypothesize that orders with low discount and “problematic” weight ranges translate to particular shipping contract specifications that reflect a lower priority by the shipment providers. Therefore, including this sort of data would further help improve the results.
Nevertheless, we believe that our current findings already suffice to address the presented problem. The models used correctly address the matter of identifying potential delays, specially when considering the patterns in our confusion matrix and classification report. Since the model tends to rarely miss the indication of actual delays, most problematic shippings will be successfully flagged to the management.