What the app does

The app shows visualizations from E-Commerce shipping data from Kaggle here. It's also interactive in creating a scatter plot based on a user-chosen feature when compared with the target variable Reached_on_time.

Feature Variables

The dataset contains 10,999 observations and 10 features. The 10 features are:

• Warehouse_block – The Company has a big Warehouse which is divided into blocks A, B, C, D, and E.

• Mode_of_Shipment – The Company ships the products in multiple ways such as Ship, Flight and Road.

• Customer_care_calls – The number of calls made from enquiry for enquiry of the shipment.

• Customer_rating – The company has rated from every customer. 1 is the lowest (Worst), 5 is the highest (Best).

• Cost_of_the_Product – Cost of the Product in US Dollars.

• Prior_purchases – The Number of Prior Purchases.

• Product_importance – The company has categorized the product in the various parameter such as low, medium, high.

• Gender – Customer's gender, either Male or Female.

• Discount_offered – Discount offered on that specific product (percentage).

• Weight_in_gms – The product weight in grams.

Target Variable

• Reached_on_time – It is the target variable, where 1 Indicates that the product has NOT reached on time and 0 indicates it has reached on time. Switched to 1 = Reached on time and 0 = Not reached on time.

Tools

This dashboard was built using Dash by Plotly. The predictor uses a Gradient Boosting Classifier model from Sci-kit Learn's library that I trained in /notebook/shipping.ipynb of this repository. This application is being served on AWS via Heroku.

Findings

Permutation Feature Importance

The top 5 most influential features in increasing model score are:

1. Weight_in_grams – 6.26%
2. Discount_offered – 2.12%
3. Prior_purchases – 1.08%
4. Cost_of_the_Product – 0.85%
5. Customer_care_calls – 0.35%

Bar Chart

Weight_in_gms

Number of products whose weight is below 2000 grams – 3241

Number of products whose weight is between 2000 and 4000 grams – 1792

Number of products whose weight is greater than 4000 grams – 5966

Discount_offered

Number of products that have a discount less than or equal to 10% – 8352

Number of products that have a discount greater than 10% – 2647

Prior Purchases – Among all orders, customers with 3 prior purchases is the most occuring (3955 orders). Distribution is positively skewed with a range between 2 and 10 prior purchases.

Customer_care_calls – Among all orders, 4 customer care calls is most common. Distribution is Normally distributed.

Cost_of_the_Product – Product costs range from $96 to $310. Looks like a blend of normal and uniform distribution.

Scatter Plot

1. Weight_in_gms – As you can see from the graph, products with weights between 2000 and 4000 grams (4.4 to 8.8 lbs) have a much higher probability of arriving late. With additional EDA, I was able to validate the accuracy of these probabilities by verifying that 99.8% of products between 2000 and 4000 grams arrive late in this dataset. I juxtaposed this with products whose weights are less than 2000 grams and products whose weights are greater than 4000 grams, and found that 67.8% and 43.2% of those products are arriving late, respectively. It's interesting to note that 67.5% of products weighing between 2000 and 4000 grams are shipped via Ship, and that 67.9% of late shipments are shipped via Ship.
2. Discount_offered – From the graph, one can see that products whose discount is above 10% see a 100% late shipment probability. This was verified with additional EDA where products with less than or equal to 10% discount and products with greater than 10% discount see a 46.9% and 100% late shipment rate, respectively.
3. Prior_purchases, Customer_care_calls, and Cost_of_the_Product do not seem to have any significant effect on late shipment probabilities.

Pie Chart

Warehouse_block – 33.3% of products come from Warehouse_block F. Every other Warehouse_block (A, B, C, D) each have 16.7% of all products in the data.

Mode_of_Shipment – Ship is the most common Mode_of_Shipment (67.8%). Flight and Road account for 16.2% and 16% of the rest of the products.

Product_importance – Products have a Product_importance of either 'low', 'medium', or 'high', which account for 48.2%, 43.2%, and 8.62% of all products, respectively.

Gender – Customers are either male (49.6%) or female (50.4%).

Reached_on_time – Of all orders in the dataset, 59.7% of orders arrived late and 40.3% arrive on time.

Improvements

In order to improve this dashboard, I would first work on improving the model. Although I delivered the model after a round of tuning using Sci-kit Learn's GridSearch, the model is currently held back by noise from a few of the existing features in the dataset. Removing these features would improve the model. For the purpose of this project, I decided to leave most of the features in training the model for the sake of interactivity.

I would also improve the UI/UX. I acknowledge this isn't the best looking dashboard but is an MVP that would be delivered to a stakeholder such as a supply chain/logistics manager. Improvements can definitely be made, and more visualizations would be great for a more comprehensive analysis of the shipping data. Thanks for taking the time to explore!