Wish Sales Projection: Project Overview

• Created models to predict the sales of items listed on the eCommerce site Wish given listed items and factors pertaining to the sales
• Performed feature engineering on various columns to help the model to perform better
• Optimized K Nearest Neighbor, Support Vector Machines, Decision Trees, Random Forest, and XGboost using GridSearchCV to reach the best model
• Utilized an image recognition API to process image data of listed items and add additional information

Code and Resources Used

Python Version: 3.7.6

Packages: pandas, numpy, scikit-learn, matplotlib, seaborn

Kaggle Dataset: https://www.kaggle.com/jmmvutu/summer-products-and-sales-in-ecommerce-wish

Image Categorization API: https://imagga.com/auto-categorization-demo

Data Cleaning

After acquiring the data, I needed to clean it so that it was usable for the model. I made the following changes:

• Added column with the number of other listings the merchant has in the data
• Cleaned various NaN and null entries within the dataset
• Reduced dimensionality of color and size columns
• Removed interdependent variables
• Processed tag variable using one hot encoding

Exploratory Data Analysis

Model Building

A train-test split was performed on the dataset with a test size of 20%. Furthermore, k-fold cross validation was utilized as a means of estimating the in-sample accuracy with k = 10.

At this stage, performance was evaluated based on accuracy.

Five machine learning algorithms were considered for this data including:

• K Nearest Neighbor - Utilized as the dataset was not extremely large and thus computationally expensive
• Support Vector Machine - Considered due to needing a classifier
• Decision Trees - Used due to the class nature of many independent and dependent variables
• Random Forest - Ensemble method for decision trees
• XGBoost - Ensemble method more optimized for performance

Model Performance

Initial Model Accuracy

Using default hyperparameters, I tested the various models considered.

• K Nearest Neighbor: Train Accuracy = 46%
• SVM with Linear Kernel: Train Accuracy = 40.86%
• Decision Trees: Train Accuracy = 48.25%
• Random Forest: Train Accuracy = 49.87%
• XGBoost: Train Accuracy = 46.98%

The Random Forest performed the best of the all the models tested.

Hyperparameter Optimization

After determining the best model to be Random Forest classifier, I used GridSearchCV to tune the hyperparameters of the model.

I considered max depth, max features, min samples leaf, min samples split, and n estimators as the hyperparameters to be optimized.

Following GridSearchCV, the model had an accuracy of 52.70%.