Product-Recsys applink
Using the customer orders and rating data from Olist E-Commerce Public Dataset, which has information about 100k orders made at multiple marketplaces in Brazil from 2016 to 2018, I trained 3 models that generate product recommendations.
NOTE
- The granularity of orders in the dataset is at product category level, thus recommendations are product categories in the true sense.
- The application is deployed on a free server that goes into sleep mode, expect a ~10 seconds delay in loading, If it takes more than 10 seconds, try refreshing! :)
Architecture Diagram
The website is divided into 3 sections.
Demographic Filtering - Recommends highest rated products in the dataset.
Method - Uses IMDB weighted average formula and recommends highest rated products.
Item based collaborative filtering - Takes user input of one product and recommends 5 similar products.
Method - Computes cosine similarity between selected product and others based on historical ratings given by customers using KNN Basic algorithm.
User based collaborative filtering - Takes user input & rating of one product and recommends 3 products based on the rating given for the selected product.
Method - Approximates the user's taste by running two algorithms in the backend, first computes the most similar user who has sufficient historical data, and second, makes recommendations based on their taste.
The model training part of this project is in the product-recsys-training repository.
This repo is currently in a development state, I'm working on setting up an easy method to run this application on local machines, will update this section soon. In the meantime check out the product-recsys-training repository.