Food-Sales-Predictions

Author

Samar Krimi

Business Problem

A sales prediction for food items sold at various stores to help the retailer to understand the properties of products and outlets that play crucial roles in increasing sales.

Data Source : sales_predictions_2023.csv

https://drive.google.com/file/d/1syH81TVrbBsdymLT_jl2JIf6IjPXtSQw/view

Data Dictionary

Cleaning and Exploring the data

8523 rows and 12 columns

Explanatory Data Visualization

• Histogram to view the distributions of various features

  

The Item_Outlet_Sales Distribution dramatically decreases with greater prices

• Heatmap of the correlation between features

  

Perfect correlation equal to 1 for : Item_Weight, Item_Visibility , Item_MRP, Outlet_Establishment_Year & Item_Outlet_Sales. Moderate correlation equal to 0.57 for : Item_MRP with Item_Outlet_Sales Very low correlation for : The rest of Items and Outlet_Establishment_Year

• Trends by Bar Chart

  

The Average outlet sales by product category show that : Seafood & Starchy Food are the best-selling they are rich in carbohydrates, nutrients, fiber and vitamins, they are a real source of energy for the body. They are ideal for physical and mental activities. In second order are Fruits and Vegetables & Snack Foods which are the most consumed by most people and thirdly Dairy, Canned & Breads. Baking Goods & Soft Drinks are the least sold, may be retailer must change their places in the store and I think that they are not recommended for people with certain diseases.

• Trends by Line Graph

The Item_MR trend seems very fluctuating over the year in which store was established. The the list price of the product increase between 2000 & 2004 years, it falls sharply in 2008 maybe due to the financial crisis. Item_Outlet_Sales trend seems almost stable through these years but it falls very sharply in 1998, maybe some sales of certains products are out of stock or a particular store had an incident in this year.

Preprocessing for Machine Learning

Preventing data leakage

Models Evaluation and Results

Linear Regression Model :

• This model performs poorly on the training set and testing set regarding R2 metric, this is a case of high bias (underfit).
• For RMSE metric, our model is incorrect by about 1,115 thousand dollars.

Regression Tree Model :

• This model performs perfect on the training data regarding R2 metric.
• It performs poorly on the testing data, this is a case of high variance (overfitting).
• For RMSE metric, our model is incorrect by about 1,5 thousand dollars it penalizes more larger errors than Linear model.

Regression Tree Model Tuned :

• The best hyperparameter is for depth=5.
• The R2 testing score become higher, it grows to 0.595 after tuning.
• The training (0.604) and test (0.595) results have moved closer to each other.
• The RMSE score is reduced, this model is incorrect by an average of 1,07 thousand dollars. By tunning the decision tree model we have reduced overfitting and RMSE metric

Recommendations for stakeholders

Item_Outlet_Sales can be improved and must be studied according to certain criteria :

• Increasing the amount of best sellers like seafood.
• Focused on fruits and vegetables beneficial to health by making special offers.
• Decrease the amount of unhealthy items like baking Goods & soft Drinks.
• Have always available items sold at various stores to prevent to avoid stock shortage especially in period of health crisis or eventual war.
• Consider selling online and on social networks.
• Changing the packaging and place in the supermarket.

Future Directions

May be change model evaluation such as an improved Random Tree Model and have a better metric results.

--------------------Project 1 - Revisited------------------------

• For Linear Regression Model, the top 3 most impactful features are:

  -Outlet_Type_Supermarket Type3 increased item outlet sales by 2,126.20

  -Item_MRP increased item outlet sales by 965.13

  -Outlet_Type_Supermarket Type1 increased item outlet sales by 737.28

  

• For Decision Tree & Random Forest Models, the top 5 most important features :

  -Item_MRP

  -Outlet_Type_Grocery_Store

  -Item_Visibility

  -Outlet_Type_Supermarket Type3

  -Item_Weight

Item_MRP is by far the single most important feature for predicting item outlet sales.

Outlet_Type_Grocery Store is the second most important.

Item_Visibility, Outlet_Type_Supermarket Type3, Item_Weight, are somewhat important.

Everything else is unimportant.

• Comparing the top 5 most important features according to SHAP vs. my RandomForest Regressor:

  • The top 5 most important features were the same according to Shap vs the RandomForest.
  • However, the order was slightly different (Item_Visibility is 4th according to shap instead of 3rd).

• The top 3 most important features and their effects were the following:

  Item_MRP: The higher the MRP, the higher the predicted sales. Outlet_Type_Grocery Store: Being a grocery store dramatically decreased the predicted sales. Outlet_Type_Supermarket Type3: Being a supermarket type 3 dramatically increased the predicted sales.

  

• The top 5 most important features according the Shap Summary Plot are :

  • Item_MRP
  • Outlet_Type_Grocery_Store
  • Outlet_Type_Supermarket Type3
  • Item_Visibility
  • Item_Weight

• Use the top features from SHAP/feature importance

  • Using top features to select 2 examples outlets :
    • Outlet_Type_Grocery Store : a store having low sales
    • Outlet_Type_Supermarket Type3 : a store having high sales
  • Create filters for each of the outlet types.
  • Select Item_Visibility within each OutletType.

• Individual Shap Force Plots

As we can see in the force plot above for example 1:

There are some features increasing the prediction such as :

- Item_Visibility
- Item_Type_Soft Drinks
- Item_Weight
- Item_MRP

Yet, there are many more features pushing the prediction in the opposite direction such as Outlet_Type_Grocery Store.

As we can see in the force plot above for example 2: While there were a bit factors decreasing the predictions. There were many more features raising the prediction such as:

- Item_Type_Frozen Foods
- Item_Visibility
- Item_MRP
- Outlet_Type_Grocery Store
- Outlet_Type_Supermarket Type3

• A Lime tabular explanation

As we can see in the LIME explanation above for example 1, there were 2 factors contributing to the raising predicted sales, such as:
    
- Item_MRP
- Item_Type_Hard Drinks
- Outlet_Type_Supermarket Type2

The other features reduced the predicted sales.

As we can see in the LIME explanation above for example 2, there were 3 factors contributing to the decreasing predicted sales, such as:
- Outlet_Type_Grocery Store
- Outlet_Type_Supermarket Type3
- Item_MRP
- Item_Type_Starchy Foods

The other features reduced the predicted sales.